Hugging the Face Chief Scientific Director is worried AI becomes ‘yes-males on servers’

The founders of the AI ​​company have a reputation to make bold statements about the potential of the technology to transform the fields, more special to the sciences. But Thomas Wolf, co-founder of Hugging Face and Chief Scientific Officer, has a more changing view.

In Essay published on x On Thursday, Wolf said he was afraid that he would become a “male on servers” there was no breakthrough in the AI ​​Research. He has developed that the current Paradigms for AI development will not give AI capable of solving outside the box, solving creativity-the vision of the problem that wins the Nobel Prizes.

“The main mistake people usually make is to think (people like) Newton or Einstein have just been a scaled good students that genius comes to life when you are a linear extrapolate a top 10%student,” Wolf wrote. “To create Einstein at a data center, we don’t just need a system that knows all the answers, but rather one that can ask questions that no one else has thought of or dared to ask.”

Wolf’s allegations are in contrast to those from Openai Executive Director Sam Altman, who in Essay earlier this year He said “Superintelligent” AI could “massively accelerate the scientific discovery”. In the same way, the anthropic CEO Dario Amade has foreseen that AI can be able to Help formulate medicines for most cancersS

Wolf’s problem with AI today – and where he thinks the technology is directed – is that it does not generate new knowledge by connecting more unrelated facts. Even with the bigger part of the Internet at its disposal, AI, as we currently understand it, fill it most in the gaps between what people already know, Wolf said.

Some AI experts including Former Google Engineer Francois Cholethave expressed similar views, arguing that although AI is capable of memorizing models of reasoning, it is unlikely to generate “new reflections” based on new situations.

Wolf believes that AI laboratories are building what are essentially “very obedient students” – not scientific revolutionaries in any sense of the phrase. AI is not encouraged today to question and offer ideas that potentially contradict his training data, he said, limiting him to answering some questions.

“To create Einstein at a data center, we don’t just need a system that knows all the answers, but rather one that can ask questions that no one else has thought or dared to ask,” Wolf said. “The one who writes,” What if everyone is wrong about it? “When all textbooks, experts and general knowledge suggest otherwise. “

Wolf thinks “Evaluation crisis“In AI, he is partly guilty of this disappointing state of affairs. It indicates the indicators that are commonly used to measure the AI ​​system improvements, most of which consist of questions that have clear, obvious and “closed” answers.

As a solution, Wolf proposes the industry of AI to “move on to a measure of knowledge and reasoning”, which is able to clarify whether AI can take “bold counter -factic approaches”, make common proposals based on “small hints”, and ask “invalid questions” that lead to “new research paths”.

The number will be to understand what this measure looks like, Wolf admits. But he thinks it can be worth the effort.

“(T) He is the most important aspect of science (e) the ability to ask the right questions and to dispute even what a person has learned,” Wolf said. “We do not need a student A+ (AI) who can answer any question with general knowledge. We need a B student who sees and questioned what everyone else has missed. “