Powerful breakthrough in artificial intelligence poised to transform world, experts say

THE artificial intelligence Generative Science (AI) has positioned itself as one of the areas of technology who showed faster progress over the last few years.

There are currently different systems that allow from writing texts to creating images, videos or music, among many other applications .

However, most of these ways of using generative AI rely on an underlying technology called a “transformer” that has considerably broader applications.

This is what the chronicler of the The Wall Street Journal and author of the book Arrive today (Harper Business, 2021), Christopher Mims, who recently wrote a text in which He discussed advances in AI .

The specialist stated that The first time Transformers were mentioned was in a 2017 article. prepared by Google researchers.

There they were described, Mims said, as “a type of AI algorithm that allows computers to understand the underlying structure of any set of data (whether words, motor data, or the amino acids of a protein) to that they can generate their own similar result “.

The transformer has enabled companies like OpenAI to develop products like ChatGPT one of the most famous systems in generative AI.

However, there are many others: from companies that offer robot taxis to others that use this technology to design new protein molecules .

The 2017 article, titled Attention is all you need, was cited more than 140,000 times in other scientific and academic articles.

According to the chronicler of Newspaper“Modern AI has long been effective at recognizing patterns in information, but Previous approaches impose serious limitations ” about what else he could do.

“With language, for example, most AI systems could only process words one by one and only evaluate them in the order they were read which limited his ability to understand what those words meant.

Mims pointed out that the Google researchers who wrote the aforementioned article They focused their experience on the language translation process .

“They realized that an AI system capable of digesting all the words in a text and giving more importance to the meaning of some words than others (in other words, read in context) could make for much better translations “.

The AI ​​researcher from Allen Institute for Artificial Intelligence Tim Dettmers said Newspaper that the importance of transformers is that they work by discovering how each piece of information absorbed by the system relates to every other piece of information it received .

According to Mims, “this level of contextual understanding allows transformer-based AI systems not only recognize patterns, but also predict what might happen next and thus generate your own new insights “.

“This ability can be extended to other data in addition to words,” he added.

For his part, the co-founder and director of EvolutionaryScale, Alexander Rives, summarized: “In a sense, patterns reveal latent structure in data “.

The aforementioned company trains its AI with the published sequences of every protein its researchers can get their hands on.

EvolutionaryScale specialists They hope that their system can, with this data and without the help of human engineers, determine the relationship between a sequence of molecular building blocks and the function of the protein it creates.

According to the rescues Newspaper, the company has already created a proof-of-concept molecule .

Mims explained that it is “a protein that functions like the one that makes jellyfish glow, but its AI-invented sequence is radically different to everything that nature has not yet invented.

“The company’s ultimate goal is to enable all types of companies, from pharmaceutical manufacturers producing new drugs to synthetic chemistry companies working on new enzymes, to create substances that would be impossible without their technology “.

“It could include bacteria equipped with new enzymes capable of digesting plastic or new drugs adapted to particular cancers of the people,” declared the columnist.

Karol Hausman, co-founder of Physical Intelligence, said his company’s goal is to “build a model that can control any robot to perform any task including all robots that exist today and those that have not yet been developed.

During a recent demonstration, the company used robotic arms powered by its AI system to fold clothes .

Although the latter may seem simple to humans, it is a very complex task for robots .

“Clothes can take any shape and requires surprising flexibility and dexterity to manipulate it so roboticists cannot program the sequence of actions that will tell a robot exactly how to move its limbs to retrieve and fold laundry,” Mims explained.

However, the mechanism of Physical Intelligence »can take clothes out of a dryer and fold them neatly using a system that has learned to do this task on its own without any human input beyond a mountain of data to digest,” the columnist said.

This has attracted the attention not only of researchers, but also of investors like Jeff Bezos and Open AI .

At the beginning of November, the company raised around $400 million .

Other companies working with automobiles have focused on discovery how to use transformer-based “visual language models” that can absorb and connect not only language, but also images .

One of them is the autonomous vehicle company Waymo. owned by Alphabet. There is also Nuro, based in California, and Wayve, based in London.

“This marks a departure from pre-Transformer autonomous driving approaches which used a combination of human-written instructions and older types of AI to process sensor data to identify objects on the road,” Mims said.

According to the chronicler, “The new transformer-based models are, in essence, a shortcut to giving autonomous driving systems the kind of general knowledge about the world that was previously very difficult to offer them. » .

“For example, in a recent paper, Waymo researchers showed that using Google’s commercial AI, called Gemini, its self-driving system could have the ability to identify and yield to objects for which it has not been trained, such as a dog crossing the street », mentioned the expert.

Although Mims said progress in the field of AI was promising, he quoted Dettmers as saying that these systems They have limits and remain unpredictable, so they will not be able to completely replace the work of humans. .

“For example, the AI ​​at the heart of EvolutionaryScale can suggest new molecules that humans can test in the laboratory, but humans have yet to synthesize and test them “, declared the columnist.

He further emphasized that transformer-based models “are far from being sufficiently reliable to fully assume conduct.”

From the vision of the co-founder of Physical Intelligence, Even though they are “setting expectations,” they are still in the early stages. in the development of certain mechanisms.