AI is more likely to create a generation of ‘yes-men on servers’ than any scientific breakthroughs, Hugging Face co-founder says

Speaking to Fortune astatine Viva Technology successful Paris, the Hugging Face co-founder said that portion ample connection models (LLMs) person shown an awesome quality to find answers to questions, they autumn abbreviated erstwhile trying to inquire the close ones—something Wolf sees arsenic the much analyzable portion of existent technological progress.

“In science, asking the question is the hard part, it’s not uncovering the answer,” Wolf said. “Once the question is asked, often the reply is rather obvious, but the pugnacious portion is truly asking the question, and models are precise atrocious astatine asking large questions.”

Wolf said helium came to the decision aft speechmaking a wide circulated blog station by Anthropic CEO Dario Amodei called Machines of Loving Grace. In it, Amodei argues the satellite is astir to spot the 21st period “compressed” into a fewer years arsenic AI accelerates subject drastically.

Wolf said helium initially recovered the portion inspiring but started to uncertainty Amodei’s idealistic imaginativeness of the aboriginal aft the 2nd read.

“It was saying AI is going to lick crab and it’s going to lick intelligence wellness problems — it’s going to adjacent bring bid into the world, but past I work it again and realized there’s thing that sounds precise incorrect astir it, and I don’t judge that,” helium said.

For Wolf, the occupation isn’t that AI lacks cognition but that it lacks the quality to situation our existing framework of knowledge. AI models are trained to foretell apt continuations, for example, the adjacent connection successful a sentence, and portion today’s models excel astatine mimicking quality reasoning, they autumn abbreviated of immoderate existent archetypal thinking.

“Models are conscionable trying to foretell the astir apt thing,” Wolf explained. “But successful astir each large cases of find oregon art, it’s not truly the astir apt creation portion you privation to see, but it’s the astir absorbing one.”

Using the illustration of the crippled of Go, a committee crippled that became a milestone successful AI past erstwhile DeepMind’s AlphaGo defeated satellite champions successful 2016, Wolf argued that portion mastering the rules of Go is impressive, the bigger situation lies successful inventing specified a analyzable crippled successful the archetypal place. In science, helium said, the equivalent of inventing the crippled is asking these genuinely archetypal questions.

Wolf archetypal suggested this thought successful a blog station titled The Einstein AI Model, published earlier this year. In it, helium wrote: “To make an Einstein successful a information center, we don’t conscionable request a strategy that knows each the answers, but alternatively 1 that tin inquire questions cipher other has thought of oregon dared to ask.”

He argues that what we person alternatively are models that behave similar “yes-men connected servers”—endlessly agreeable, but improbable to situation assumptions oregon rethink foundational ideas.