Hugging Face’s chief science officer worries AI is becoming ‘yes-men on servers’

The founders of the AI ​​Company have a reputation for claiming the fields of re -shape the technology, especially the possibilities of science. However, the hug face co-founder and Chief Science Officer Thomas Wolf has been further measured.

A article On Thursday, Wolf said he feared “yes-men” on AI AI research missing “AI” servers. He detailed that the current AI development examples would not yield the AI ​​to solve the creative problem-which the Nobel Prize was won.

“The main mistake of people who usually make [people like] Wolf writes that Newton or Einstein just had a good student to scald -up, when you extrostle a top -10% of the students, a talent is vivid. “” To create Einstein at a data center, we don’t only need a system that knows all the answers, but also asking a question that no one else can think or asks. “

Olph’s speeches opposite OpenAI CEO Sam Altman, who was one who was Articles at the beginning of this year It was said that “superintendent” AI “could greatly accelerate the scientific discovery.” Similarly, the ethnographic chief executive officer Dario Amodai predicts the AI ​​cans Help in creating cure for most types of cancerThe

Today is the problem of Olf with AI – and where he thinks the technology is heading – it is that it does not create any new knowledge by attaching the previously related information. Even if most internet disposes, AI is currently understanding that it is mostly in the case that people already fill in the gap between what he knows, Olf said.

Some AI experts including Former Google Engineer Francois ChollettSimilar views, by arguing that AI can be able to memorize reasonable patterns, is less likely to generate “new logic” based on the situation in the novel.

Wolf thinks that AI labs are basically “very obedient students” that are creating – not scientific revolutionary in any sense of these phrases. AI was not encouraged to question and proposes today, which is possible to probably go against its training information, he said it limited to answer the familiar questions.

“To create Einstein at the Data Center, we only need a system that knows all the answers, but can ask a question that no one else does or dare to ask,” Olph said. “Anyone who wrote ‘what would happen if everyone was wrong about it?’ When all the textbooks, experts and common sense otherwise suggests ”” “

Olph thinks that “Crisis“AI is partially responsible for this disappointing situation. He usually points to the criteria used to measure the improvements of the AI ​​system, most of which are made up of questions that are clear, obvious and “close” answers.

As a solution, Wolf suggested that AI industry was able to explain whether “AI” which AI was able to take “bold counter factorial approaches”, creating general proposals on the basis of “small hints” and asking “non-Buddhist questions” that leads to the path of new research. “

The strategy will determine what this measure looks like, Olf acknowledges. However, he thinks it may be suitable for the effort.

“[T]He is the most important aspect of science [is] Skill to ask the right questions and even someone who has learned what has learned, “said Olf.“ We don’t need an A+ [AI] The student who can answer each question with common sense. We need a B student that everyone else sees and asks and asks. “