According to Thomas Wolf, top scientist from Face, it is unlikely that the current AI systems will make the scientific discoveries that some leading laboratories hope for.
Speak to Assets In the case of Viva technology in Paris, the co-founder of the hugging face said that large language models (LLMS) have shown an impressive ability to find answers to questions, when trying to ask the right ones, too short, the wolf sees as a more complex part of real scientific progress.
“In science it is the difficult part not to find the answer,” said Wolf, “said Wolf.” As soon as the question has been asked, the answer is often quite obvious, but the difficult part is really to ask the question, and models are very bad to ask great questions. ”
Wolf said Machines of loving grace. In it, Amodei argues that the world will be “compressed” for the 21st century in a few years, while AI drastically accelerates science.
Wolf said that he initially found the play inspiring, but after the second reading of Amodeis Irranged idealistic vision of the future.
“It was said that Ai will solve cancer and that it will solve mental health problems – it will even bring peace to the world, but then I read it again and realized that something that sounds very wrong, and I don’t think so,” he said.
For Wolf, the problem is not that AI has no knowledge, but that there is no ability to challenge our existing knowledge framework. AI models are trained in order to predict the likely continuations, for example the next word in one sentence, and while today’s models are characterized in imitation of human thinking, they fall beyond real original thinking.
“Models only try to predict the most likely,” said Wolf. “But in almost all major cases of discovery or art, it is not really the most likely feat that you want to see, but it’s the most interesting.”
With the example of the Game of Go, a board game that became a milestone in AI history, when Deepmind’s Alphago defeated world champion in 2016, Wolf argued that the mastery of the rules of Go is impressive, the greater challenge in the invention of such a complex game. In science, he said, the equivalent to invent the game is these really original questions.
Wolf first suggested this idea in a blog post with the title with the title The Einstein AI modelPublished at the beginning of this year. In it he wrote: “To create an Einstein in a data center, we not only need a system that knows all the answers, but also one that can ask questions that nobody thought of or has dared to ask.”
He argues that instead we are models that are behavioral-endlessly pleasant like “Ja-Men on servers”, but it is unlikely that you question assumptions or rethink basic ideas.
