Hugging Face cofounder Thomas Wolf has just challenged Anthropic CEO’s vision for AI’s future –

Thomas Wolfis the cofounder of AI company Hugging Facehas issued a challenge to the most optimistic visions in the tech industry of artificial intelligence. It argues that today’s AI is fundamentally incapable of delivering on the scientific revolutions its creators promised.

In a provocative In a blog post posted on his website this morning by Wolf, he directly challenges the widely circulated vision that Anthropic CEO Dario Amedei had, who predicted advanced AI would lead to a ” Compressed 21st Century“where decades of scientific advancement could unfold in only a few years.

Wolf writes, “I’m afraid AI will not give us a “compressed 21st Century,”” arguing that current AI is more likely to produce ” A country of yes-men in servers” rather than ” Amodei imagines a “country of geniuses”that Amodei describes.

This exchange highlights the growing divide between AI leaders’ views on how AI can transform scientific discovery and problem solving, with major implications for policy decisions, business strategies and research priorities.

Why academic excellence does not equal scientific genius

Wolf bases his critique on personal experience. Despite being a straight A student at MIT, he describes finding out that he was a mediocre researcher when he started his PhD. This experience led him to believe that academic success and genius in science require fundamentally different mental attitudes — the first rewarding conformity and the latter demanding rebelliousness against established thinking. Wolf explains that the main mistake people make is to think of Newton and Einstein as just scaled-up, good students. “A real scientific breakthrough is Copernicus’ proposal, against all knowledge of his day — in ML terminology, we would say “despite all his data set” — that the Earth may orbit the Sun rather than the opposite. Machines of Loving graceessay presents a radically new perspective. He describes a world where AI could achieve a century of progress in neuroscience, biology and other fields by operating at “10x to 100x human speed”.

Amodei envisions “reliable treatment and prevention of almost all infectious diseases”, “elimination” of most cancers, effective cures for genetic diseases, and possibly doubling the human lifespan. All of this is accelerated by AI. “I believe that intelligence is a major factor in these discoveries and that everything else follows from them,” writes Amodei.

Are we testing AI on conformity rather than creativity? The benchmark problem preventing scientific discovery

Wolf’s critique highlights a reality that is often overlooked in AI development. Our benchmarks are designed to measure convergent rather than divergent thought. Current AI systems are good at producing answers that conform to existing knowledge consensus but struggle to produce the kind of contrarian and paradigm-challenging insight that drives scientific revolutions.

Industry has heavily invested in measuring how well AI can answer questions that have established answers, solve known problems, and fit into existing frameworks of knowledge. This biases systems to conform rather than challenge.

Wolf criticizes current AI benchmarks such as ” Humanity’s last exam” and ” Frontier Mathis a test of AI systems that tests their ability to generate new hypotheses and challenge existing paradigms, rather than their ability in generating innovative hypotheses.

Wolf writes that these benchmarks are used to test whether AI models can answer questions we already have the answers to. “However, the real scientific breakthroughs won’t come from answering questions we already know the answer to, but by asking new questions and challenging common conceptions and prior ideas.”

The critique points out a deeper issue with how we conceptualize AI. The current focus on parameter counts, training data volumes, and benchmark performances may create AI equivalents of excellent students instead of revolutionary thinkers.

Billions on the line: How the debate between obedient students and revolutionaries will shape AI investment strategies

The intellectual divide has significant implications for the AI industry as well as the broader business eco-system.

Companies that align with Amodei’s vision may prioritize scaling AI systems up to unprecedented sizes. They expect discontinuous innovation from increased computational power, and broader knowledge integration. This approach is at the core of the strategies of companies like Anthropic Openai as well as other frontier AI laboratories that have collectively raised In recent years, tens[of billions]of dollars have been spent on AI systems.

Wolf’s perspective suggests, on the other hand, that greater returns could come from developing AI system specifically designed to challenge current knowledge, explore counterfactuals, and generate novel hypothesis — capabilities not necessarily arising from current training methods.

Wolf explains that “we’re currently building very compliant students, not revolutionary.” This is ideal for today’s primary goal of creating great assistants, and overly compliant assisters. But until we find ways to encourage them to question their current knowledge and suggest ideas that could go against previous training data, they will not give us scientific breakthroughs yet.

This debate raises important strategic questions for enterprise leaders who are betting on AI to fuel innovation. If Wolf is right, organizations that invest in current AI systems expecting revolutionary scientific breakthroughs might need to temper their expectations. The real value could be found in incremental improvements to current processes or in collaborative human-AI approaches, where humans provide the paradigm challenging intuitions and AI systems do the computational heavy lifting.

The $184 billion dollar question: Is AI prepared to deliver on its promises scientifically? This exchange occurs at a pivotal point in the evolution of the AI industry. After years of explosive growth and investment in AI capabilities, both public and privately-held stakeholders are increasingly focused upon the practical benefits from these technologies. Recent data from venture-capital analytics firm PitchBook show AI funding has reached $130 billion worldwide in 2024with healthcare and scientific discoveries applications attracting special interest. Questions about the tangible scientific breakthroughs that can be expected from these investments are becoming more and more frequent.

This debate between Wolf and Amodei represents a deeper philosophical division in AI development, which has been simmering under the surface of industry discussion. On one side are the scaling optimists who believe that continual improvements in model size and data volume, as well as training techniques, will eventually lead to systems that can provide revolutionary insights. On the other hand, architecture skeptics argue that current systems may not be able to make the cognitive leaps necessary for scientific revolutions. This debate is particularly important because it involves two respected leaders in AI development. Both cannot be dismissed as being uninformed or resistant of technological progress.

Beyond scaling: How AI of tomorrow might need to think like scientific rebels.

This tension between these perspectives suggests a possible evolution in the way AI systems are designed. Wolf’s critique does not suggest abandoning existing approaches, but rather enhancing them with new metrics and techniques that are specifically designed to foster contrarian thinking.

Wolf’s post suggests that new benchmarks be developed to test if scientific AI models are able to “challenge the knowledge they have gained from their training data” and “take counterfactual approaches.” It is not a call for less AI investment but rather for more thoughtful investments that consider the full spectrum cognitive capabilities required for scientific progress. This nuanced approach acknowledges AI’s enormous potential, while acknowledging that current systems may excel in certain types of intelligence and struggle with others. The way forward is likely to involve developing complementary approaches which leverage the strengths of existing systems while finding a way to address their shortcomings. The implications for businesses and research institutions that are navigating AI strategies are significant. Organisations may have to develop evaluation frameworks for assessing not only how well AI systems can answer existing questions but also how effectively they can generate new ones. They may need models of collaboration between humans and AI that combine the pattern-matching, computational abilities of AI and the paradigm-challenging intuitiveness of human experts.

Finding the middle way: How AI can combine computational power and revolutionary thinking

The most valuable outcome from this exchange may be that it pushes industry towards a more balanced understanding both of AI’s potential as well as its limitations. Amodei’s vision is a powerful reminder of the impact AI can have on multiple domains at once. Wolf’s critique (19459057) provides a needed counterbalance by highlighting the types of cognitive abilities required for truly revolutionary progress.

This tension between optimism, skepticism and scaling existing approaches versus developing new approaches will likely drive innovation in AI as the industry moves forward. Understanding both perspectives will help organizations develop more nuanced strategies to maximize the potential of existing systems, while also investing in solutions that address their limitations .

The question is not whether Wolf or Amodei are correct, but how their contrasting views can inform a comprehensive approach to developing AI that does more than just answer the questions we currently have, but also helps us discover questions we haven’t thought of yet.