AI Becoming 'Yes-Men on Servers', Warns Hugging Face CSO

Thomas Wolf, co-founder and chief science officer at Hugging Face, offers a more cautious perspective on AI's potential to revolutionize science, contrasting with the bold claims often made by other AI company founders. In an essay shared on X, Wolf expressed concerns that AI might become mere "yes-men on servers" without significant breakthroughs in research. He argues that the current trajectory of AI development won't produce systems capable of the creative, out-of-the-box thinking needed for groundbreaking scientific achievements, like those that earn Nobel Prizes. Wolf challenges the notion that geniuses like Newton or Einstein are simply scaled-up versions of top students. He believes that to replicate such genius in a data center, AI must not only know all the answers but also be able to pose questions that no one else has considered or dared to ask. This perspective differs from that of OpenAI CEO Sam Altman, who envisions "superintelligent" AI accelerating scientific discovery, and Anthropic CEO Dario Amodei, who predicts AI could help find cures for most cancers. Wolf's critique of current AI is that it fails to generate new knowledge by connecting previously unrelated facts. Despite having access to vast amounts of internet data, AI primarily fills in the gaps of existing human knowledge, according to Wolf. This view is echoed by other AI experts, such as former Google engineer François Chollet, who argue that while AI can memorize reasoning patterns, it struggles to generate new reasoning in novel situations. Wolf suggests that AI labs are currently building "very obedient students" rather than the scientific revolutionaries needed for true innovation. He points out that today's AI systems are not designed to question or propose ideas that contradict their training data, limiting them to answering known questions. Wolf emphasizes the need for an AI that can challenge established knowledge and ask, "What if everyone is wrong about this?" even when all evidence suggests otherwise. He attributes part of the problem to an "evaluation crisis" in AI, where benchmarks used to measure system improvements often focus on questions with clear, closed-ended answers. To address this, Wolf proposes that the AI industry should shift towards evaluating AI's ability to take bold counterfactual approaches, make general proposals from minimal hints, and ask non-obvious questions that could lead to new research paths. While acknowledging the challenge in defining such a measure, Wolf believes it's a worthwhile endeavor. He stresses that the essence of science lies in asking the right questions and challenging learned knowledge. Rather than an A+ student who can answer every question with general knowledge, Wolf argues that what's needed is a B student who can see and question what others have missed.

Related article

Comments (25)