Founders of AI companies are often known for their ambitious assertions about the technology’s capacity to transform various sectors, especially in the scientific realm. However, Thomas Wolf, co-founder and chief science officer of Hugging Face, adopts a more cautious perspective.
In an article published on X on Thursday, Wolf expressed concerns about AI evolving into “yes-men on servers” unless there’s a significant breakthrough in AI research. He argues that the current paradigms of AI development are unlikely to produce systems that can think laterally or engage in creative problem-solving similar to that which earns Nobel Prizes.
“A common misjudgment is considering figures like Newton or Einstein merely as enhanced versions of top students; genius cannot be simply extrapolated from a high achiever,” Wolf remarked. “To cultivate an Einstein in a data center, it’s essential to have a system capable not just of knowing all the answers, but also of generating questions that nobody has thought to ask.”
Wolf’s views contrast starkly with those of OpenAI’s CEO Sam Altman, who claimed in an essay earlier this year that “superintelligent” AI could significantly accelerate the pace of scientific discovery. Likewise, Anthropic CEO Dario Amodei has suggested that AI might help in discovering cures for the vast majority of cancers.
Wolf critiques the state of AI today, suggesting it fails to produce new insights by linking previously unrelated pieces of information. He asserts that even with access to a wealth of internet knowledge, current AI primarily fills in the blanks of what humans already know.
Several experts in AI, including former Google engineer Francois Chollet, share similar sentiments, contending that while AI may excel in recalling reasoning patterns, it is unlikely to create “new reasoning” in the context of novel scenarios.
Wolf believes AI laboratories are essentially training “very compliant students,” rather than innovative thinkers. Today’s AI is not compelled to question established notions, limiting it to tackling known inquiries.
“To generate an Einstein in a data center, we require not just a system that possesses all the answers, but one that can question the validity of established beliefs,” Wolf stated. “One that dares to ask, ‘What if everyone is mistaken about this?’ even when all experts, textbooks, and standard knowledge lean towards a consensus.”
Wolf attributes this disappointing scenario to an “evaluation crisis” in AI. He notes that many benchmarks for assessing AI systems are reliant on questions with straightforward, unequivocal answers.
As a potential remedy, Wolf recommends that the AI sector adopt a new metric for assessing knowledge and reasoning capabilities that can reveal whether AI can pursue “audacious counterfactual ideas,” generate general concepts from “subtle signals,” and pose “non-obvious questions” that can pave the way for “new avenues of research.”
Wolf acknowledges that identifying what this new metric should entail is a challenging task, but he believes it is a pursuit worth undertaking.
“The essence of science is the ability to pose the right questions and to challenge established knowledge,” Wolf asserted. “We don’t need a top-tier AI student who can respond accurately to every question. What we really need is a modestly performing student who can identify and question overlooked details.”
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
