Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More<\/em><\/p>\n\n\n\n Thomas Wolf, cofounder of AI company Hugging Face, has issued a stark challenge to the tech industry\u2019s most optimistic visions of artificial intelligence, arguing that today\u2019s AI systems are fundamentally incapable of delivering the scientific revolutions their creators promise.<\/p>\n\n\n\n In a provocative blog post published on his personal website this morning, Wolf directly confronts the widely circulated vision of Anthropic CEO Dario Amodei, who predicted that advanced AI would deliver a \u201ccompressed 21st century\u201d where decades of scientific progress could unfold in just years.<\/p>\n\n\n\n \u201cI\u2019m afraid AI won\u2019t give us a \u2018compressed 21st century,’\u201d Wolf writes in his post, arguing that current AI systems are more likely to produce \u201ca country of yes-men on servers\u201d rather than the \u201ccountry of geniuses\u201d that Amodei envisions.<\/p>\n\n\n\n The exchange highlights a growing divide in how AI leaders think about the technology\u2019s potential to transform scientific discovery and problem-solving, with major implications for business strategies, research priorities and policy decisions.<\/p>\n\n\n\n Wolf grounds his critique in personal experience. Despite being a straight-A student who attended MIT, he describes discovering he was a \u201cpretty average, underwhelming, mediocre researcher\u201d when he began his PhD work. This experience shaped his view that academic success and scientific genius require fundamentally different mental approaches \u2014 the former rewarding conformity, the latter demanding rebellion against established thinking.<\/p>\n\n\n\n \u201cThe main mistake people usually make is thinking Newton or Einstein were just scaled-up good students,\u201d Wolf explains. \u201cA real science breakthrough is Copernicus proposing, against all the knowledge of his days \u2014 in ML terms we would say \u2018despite all his training dataset\u2019 \u2014 that the earth may orbit the sun rather than the other way around.\u201d<\/p>\n\n\n\n Amodei\u2019s vision, published last October in his \u201cMachines of Loving Grace\u201d essay, presents a radically different perspective. He describes a future where AI, operating at \u201c10x-100x human speed\u201d and with intellect exceeding that of Nobel Prize winners, could deliver a century\u2019s worth of progress in biology, neuroscience and other fields within five to 10 years.<\/p>\n\n\n\n Amodei envisions \u201creliable prevention and treatment of nearly all natural infectious disease,\u201d \u201celimination of most cancer,\u201d effective cures for genetic disease, and potentially doubling human lifespan, all accelerated by AI. \u201cI think the returns to intelligence are high for these discoveries, and that everything else in biology and medicine mostly follows from them,\u201d he writes.<\/p>\n\n\n\n This fundamental tension in Wolf\u2019s critique reveals an often-overlooked reality in AI development: Our benchmarks are primarily designed to measure convergent thinking rather than divergent thinking. Current AI systems excel at producing answers that align with existing knowledge consensus, but struggle with the kind of contrarian, paradigm-challenging insights that drive scientific revolutions.<\/p>\n\n\n\n The industry has invested heavily in measuring how well AI systems can answer questions with established answers, solve problems with known solutions, and fit within existing frameworks of understanding. This creates a systemic bias toward systems that conform rather than challenge.<\/p>\n\n\n\n Wolf specifically critiques current AI evaluation benchmarks like \u201cHumanity\u2019s Last Exam\u201d and \u201cFrontier Math,\u201d which test AI systems on difficult questions with known answers rather than their ability to generate innovative hypotheses or challenge existing paradigms.<\/p>\n\n\n\n \u201cThese benchmarks test if AI models can find the right answers to a set of questions we already know the answer to,\u201d Wolf writes. \u201cHowever, real scientific breakthroughs will come not from answering known questions, but from asking challenging new questions and questioning common conceptions and previous ideas.\u201d<\/p>\n\n\n\n This critique points to a deeper issue in how we conceptualize artificial intelligence. The current focus on parameter count, training data volume, and benchmark performance may be creating the AI equivalent of excellent students rather than revolutionary thinkers.<\/p>\n\n\n\n This intellectual divide has substantial implications for the AI industry and the broader business ecosystem.<\/p>\n\n\n\n Companies aligning with Amodei\u2019s vision might prioritize scaling AI systems to unprecedented sizes, expecting discontinuous innovation to emerge from increased computational power and broader knowledge integration. This approach underpins the strategies of firms like Anthropic, OpenAI and other frontier AI labs that have collectively raised tens of billions of dollars in recent years.<\/p>\n\n\n\n Conversely, Wolf\u2019s perspective suggests that greater returns might come from developing AI systems specifically designed to challenge existing knowledge, explore counterfactuals and generate novel hypotheses \u2014 capabilities not necessarily emerging from current training methodologies.<\/p>\n\n\n\n \u201cWe\u2019re currently building very obedient students, not revolutionaries,\u201d Wolf explains. \u201cThis is perfect for today\u2019s main goal in the field of creating great assistants and overly compliant helpers. But until we find a way to incentivize them to question their knowledge and propose ideas that potentially go against past training data, they won\u2019t give us scientific revolutions yet.\u201d<\/p>\n\n\n\n For enterprise leaders betting on AI to drive innovation, this debate raises crucial strategic questions. If Wolf is correct, organizations investing in current AI systems with the expectation of revolutionary scientific breakthroughs may need to temper their expectations. The real value may be in more incremental improvements to existing processes, or in deploying human-AI collaborative approaches where humans provide the paradigm-challenging intuitions while AI systems handle computational heavy lifting.<\/p>\n\n\n\n This exchange comes at a pivotal moment in the AI industry\u2019s evolution. After years of explosive growth in AI capabilities and investment, both public and private stakeholders are increasingly focused on practical returns from these technologies.<\/p>\n\n\n\n Recent data from venture capital analytics firm PitchBook shows AI funding reached $130 billion globally in 2024, with healthcare and scientific discovery applications attracting particular interest. Yet questions about tangible scientific breakthroughs from these investments have grown more insistent.<\/p>\n\n\n\n The Wolf-Amodei debate represents a deeper philosophical divide in AI development that has been simmering beneath the surface of industry discussions. On one side stand the scaling optimists, who believe that continuous improvements in model size, data volume and training techniques will eventually yield systems capable of revolutionary insights. On the other side are architecture skeptics, who argue that fundamental limitations in how current systems are designed may prevent them from making the kind of cognitive leaps that characterize scientific revolutions.<\/p>\n\n\n\n What makes this debate particularly significant is that it\u2019s occurring between two respected leaders who have both been at the forefront of AI development. Neither can be dismissed as simply uninformed or resistant to technological progress.<\/p>\n\n\n\n The tension between these perspectives points to a potential evolution in how AI systems are designed and evaluated. Wolf\u2019s critique doesn\u2019t suggest abandoning current approaches, but rather augmenting them with new techniques and metrics specifically aimed at fostering contrarian thinking.<\/p>\n\n\n\n In his post, Wolf suggests that new benchmarks should be developed to test whether scientific AI models can \u201cchallenge their own training data knowledge\u201d and \u201ctake bold counterfactual approaches.\u201d This represents a call not for less AI investment, but for more thoughtful investment that considers the full spectrum of cognitive capabilities needed for scientific progress.<\/p>\n\n\n\n This nuanced view acknowledges AI\u2019s tremendous potential while recognizing that current systems may excel at particular types of intelligence while struggling with others. The path forward likely involves developing complementary approaches that leverage the strengths of current systems while finding ways to address their limitations.<\/p>\n\n\n\n For businesses and research institutions navigating AI strategy, the implications are substantial. Organizations may need to develop evaluation frameworks that assess not just how well AI systems answer existing questions, but how effectively they generate new ones. They may need to design human-AI collaboration models that pair the pattern-matching and computational abilities of AI with the paradigm-challenging intuitions of human experts.<\/p>\n\n\n\n Perhaps the most valuable outcome of this exchange is that it pushes the industry toward a more balanced understanding of both AI\u2019s potential and its limitations. Amodei\u2019s vision offers a compelling reminder of the transformative impact AI could have across multiple domains simultaneously. Wolf\u2019s critique provides a necessary counterbalance, highlighting the specific types of cognitive capabilities needed for truly revolutionary progress.<\/p>\n\n\n\n As the industry moves forward, this tension between optimism and skepticism, between scaling existing approaches and developing new ones, will likely drive the next wave of innovation in AI development. By understanding both perspectives, organizations can develop more nuanced strategies that maximize the potential of current systems while also investing in approaches that address their limitations.<\/p>\n\n\n\n For now, the question isn\u2019t whether Wolf or Amodei is correct, but rather how their contrasting visions can inform a more comprehensive approach to developing artificial intelligence that doesn\u2019t just excel at answering the questions we already have, but helps us discover the questions we haven\u2019t yet thought to ask.<\/p>\n Daily insights on business use cases with VB Daily<\/strong><\/p>\n If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI.<\/p>\n \n\t\t\t\t\tThanks for subscribing. Check out more VB newsletters here.\n\t\t\t\t<\/p>\n An error occured.<\/p>\n<\/p><\/div>\n
\n<\/div>From straight-A student to \u2018mediocre researcher\u2019: Why academic excellence doesn\u2019t equal scientific genius<\/h2>\n\n\n\n
Are we testing AI for conformity instead of creativity? The benchmark problem holding back scientific discovery<\/h2>\n\n\n\n
Billions at stake: How the \u2018obedient students vs. revolutionaries\u2019 debate will shape AI investment strategy<\/h2>\n\n\n\n
The $184 billion question: Is AI ready to deliver on its scientific promises?<\/h2>\n\n\n\n
Beyond scaling: How tomorrow\u2019s AI might need to think more like scientific rebels<\/h2>\n\n\n\n
Finding the middle path: How AI could combine computational power with revolutionary thinking<\/h2>\n\n\n\n
![\"\"\/](\"https:\/\/venturebeat.com\/wp-content\/themes\/vb-news\/brand\/img\/vb-daily-phone.png\")
\n\t\t\t\t<\/div>\n<\/p><\/div>\n<\/div>\t\t\t<\/div>\r\n