How LLMs Will Change Science
The Return of the Polymath
History's greatest scientific breakthroughs often came from minds that refused to respect the boundaries between disciplines. Leonardo da Vinci moved fluidly between art, engineering, anatomy, and astronomy. Benjamin Franklin was simultaneously a politician, writer, inventor, and scientist. Indeed, many of the greatest contributors to the early development of science were interdisciplinary thinkers: Pascal, Boyle, Leibniz, Newton, Young, etc.
These polymaths succeeded not despite their broad interests, but because of them. Their wide-ranging knowledge allowed them to spot connections that were often overlooked by specialists. Their greatest advantage wasn't specialized knowledge, but rather their ability to ask questions no one else was asking.
Yet over the past century, the polymath became an endangered species. As knowledge expanded exponentially across every discipline, specialization became less of a choice and more of a necessity. The physicist, the molecular biologist, and the evolutionary psychologist may work at the same university, but they publish in different journals, attend different conferences, and often lack even the shared vocabulary to meaningfully discuss each other's work.
This trend toward hyper-specialization made sense in a world where human cognitive limitations forced difficult tradeoffs between knowledge breadth and depth. But what if those tradeoffs no longer applied? What if a new technology could dramatically reduce the barriers to entry across all fields of knowledge?
This is precisely what's happening with the rise of large language models. These AI systems are not just search engines with better interfaces - they're dynamic thought partners that can explain complex concepts, translate between specialist vocabularies, and help navigate the vast territory of human knowledge. For the first time in generations, the pendulum is swinging back toward the generalist, creating a renaissance for cross-disciplinary thinking that may reshape how scientific discovery happens.
Knowledge Depth vs. Breadth
The specialization of modern science wasn't an arbitrary development. It was a rational response to the knowledge explosion that has occurred over the past couple of centuries. As each field deepened its understanding, mastering even a narrow subfield could require a decade or more of dedicated study. The physicist who wants to contribute to string theory must first spend years learning advanced mathematics. The neuroscientist needs intimate familiarity with complex biological systems before making meaningful contributions.
However, this necessary depth came at a cost. Knowledge silos emerged, reinforced by academic structures that rewarded narrow expertise over broad integration. Departments became more isolated, graduate programs more specialized, and journals more niche. The incentive structure of academia—publish or perish, citation metrics, tenure requirements—further entrenched specialization.
The result is a scientific landscape where breakthrough ideas often struggle to cross disciplinary boundaries. A concept from the frontiers of say materials science might help solve a persistent problem in say biology, but the two communities rarely interact. A statistical approach common in ecology might solve problems in economics, but specialists in either field have little opportunity or incentive to bridge that gap.
Even when researchers recognize the potential value of cross-disciplinary work, practical barriers loom large. Learning the foundational concepts, terminology, and methodologies of an unfamiliar field is daunting. Writing grant proposals or publishing papers outside one's specialty involves navigating unfamiliar gatekeepers and expectations. The academic system, designed to foster deep expertise, inadvertently created barriers to the kind of cross-pollination that historically drove many scientific revolutions.
The cruel irony is that as scientific challenges grow more complex, from climate change to artificial intelligence to fundamental physics, they increasingly demand interdisciplinary approaches. The very structure that has helped to advance scientific knowledge now sometimes hinders the integrated thinking needed to solve our most pressing problems.
My Personal Journey From Curious Generalist to Scientific Contributor
My own entry into scientific exploration followed an unconventional path. I was originally planning to follow a traditional degree path in physics. However, I dropped out during my undergrad to pursue an entrepreneurial path instead. Yet I never lost my love of science and continued to fill my free time following the latest developments in the scientific literature and reading dozens of books a year on a variety of topics and disciplines.
Over the past few years, I found myself with more time to indulge my curiosity and began diving deeper into fundamental physics. Unlike specialists who spend decades building expertise in a single discipline, I came to my questions as a scientifically literate generalist with no formal training, but with a decades-long habit of voracious reading across disciplines and staying up-to-date on the latest developments in the field.
This background gave me both advantages and disadvantages. I lacked the mathematical fluency and specialized knowledge that professional physicists develop through years of focused study. But I also approached questions without the preconceptions that sometimes constrain specialists.
My scientific literacy spanned popular physics and mathematics, neuroscience, biology, chemistry, engineering, philosophy, psychology, economics, logic and so on. So when I began exploring questions, say about the emergence of complexity in a universe governed by entropy or about the nature of dark energy, I didn't approach them as strictly a thermodynamics problem or a cosmology problem.
The absence of formal training in any specific discipline meant I couldn't rely on established methods or conventional wisdom. Instead, I had to build understanding from first principles, questioning fundamental assumptions along the way. In fact, I started by asking lots and lots of silly questions…but with each round of questions, I learned more and was able to continually improve the next round of questions. When I started diving deeper into the Second Law of Thermodynamics, I didn't simply accept its standard formulations; I asked whether there might be alternative interpretations consistent with observed phenomena but offering different implications.
This outsider perspective isn't always an advantage, of course. It's easy to reinvent wheels or pursue dead ends that specialists long ago discovered and abandoned for good reason. But occasionally, a fresh perspective allows you to ask questions that have been overlooked precisely because they fall between established disciplines or challenge deeply held assumptions.
What I discovered through this process wasn't merely insights about scientific questions. In some ways, it felt like I was rediscovering a methodology for exploration itself; an approach that leverages the unique advantages of the generalist mindset. With the advent of AI tools as thought partners, this approach has not only become viable once more, but could potentially be revolutionary.
How AI Tools Change the Game
The turning point in my personal scientific research projects came with the emergence of recent frontier LLMs. These AI systems transformed from mere information retrieval tools into genuine thought partners. The best frontier models became collaborators capable of explaining complex concepts, working through mathematical derivations, and connecting ideas across disparate fields.
Before these tools, my generalist approach would have hit insurmountable barriers. Learning the mathematical formalism needed to engage with modern physics would have required years of dedicated study. Understanding specialized terminology across multiple disciplines would have meant painstakingly cross-referencing textbooks and research papers. The cognitive load would have overwhelmed any individual's capacity.
Instead, I found myself able to use these AI tools to rapidly iterate on ideas. When exploring concepts typically reserved for specialists, such as say quintessence models for dark energy or Ostrogradsky instabilities in modified approaches to GR, I could immediately ask for explanations at varying levels of depth - from physically intuitive to mathematically precise. If an explanation referenced unfamiliar mathematics, I could request clarification on those specific elements, learning the necessary tools in real-time. The AI served as a translator between mathematical formalism and intuitive understanding.
The workflow that emerged was far more dynamic than traditional research. Rather than spending months mastering foundations before asking original questions, I could immediately begin exploring ideas, filling knowledge gaps as they arose. This iterative process—asking questions, receiving explanations, using that information to ask even better questions—created an accelerated learning loop that would have been impossible with traditional methods.
Perhaps most valuable was the ability to rapidly change direction when ideas didn't pan out. Traditional academic research often suffers from the sunk cost fallacy. Having invested years in developing expertise in a particular approach, researchers understandably hesitate to abandon it, even when evidence suggests it's unproductive. With AI assistance, I could explore multiple conceptual frameworks in parallel, quickly discarding those that showed fundamental flaws without the psychological burden of wasted effort.
When I developed initial hypotheses about connections between gravitational potential and dark energy, I could quickly test the mathematical consistency of these ideas without spending years learning the intricacies of tensor calculus. When these approaches revealed limitations, I could pivot to exploring new avenues, again with the AI handling the mathematical heavy lifting while I focused on the conceptual implications.
This isn't to suggest the AI did the creative work. Current AI tools excel at explaining established knowledge and working through mathematical derivations, but the questions, conceptual leaps, and novel hypotheses are still, at least for the time being, the domain of human creativity.
What emerged was a new kind of scientific exploration where the generalist could follow their curiosity across disciplinary boundaries without being stopped by specialized knowledge barriers, while still engaging with complex ideas at a substantial level of rigor.
The New Scientific Method: Question-Focused vs. Knowledge-Focused
This AI-assisted approach represents what I believe will be a real shift in how scientific exploration takes place. Traditional scientific education focuses on building deep knowledge within a field before attempting to ask original questions. The assumption is that meaningful questions can only emerge from comprehensive domain expertise.
The new paradigm will invert this relationship. It will prioritize question-asking as the primary skill, with knowledge acquisition becoming an on-demand process driven by the needs of specific inquiries. This isn't to diminish the value of deep knowledge, but rather to recognize that with AI assistance, the acquisition of relevant knowledge can happen dynamically during the exploration process rather than requiring years of front-loaded specialization.
This shift favors a different set of cognitive skills and educational approaches. Rather than rewarding memorization, technical percision, and accumulation of existing knowledge, it values creativity, conceptual flexibility, and comfort with uncertainty. It privileges minds that can make unexpected connections, challenge conventional wisdom, and approach problems from multiple perspectives.
The iterative cycle at the heart of this approach—form questions, seek information, form better questions, repeat—mirrors the core of traditional scientific discovery but dramatically accelerates the pace. Where a traditional researcher might spend years developing expertise before asking novel questions, the AI-empowered generalist can begin asking substantive questions immediately, with each cycle refining both the questions and the questioner's understanding.
This acceleration is particularly valuable for addressing conceptual blind spots in established fields. When specialists have worked within a paradigm for decades, certain assumptions may become so ingrained that they're no longer recognized as assumptions at all—they're simply "how things are." A generalist equipped with AI assistance can approach these fields with fresh eyes, questioning foundational principles that specialists might take for granted.
This approach also helps overcome psychological biases that impede scientific progress. Confirmation bias leads researchers to favor evidence supporting their existing theories. Professional incentives discourage challenging established paradigms. The institutional prestige associated with certain approaches creates inertia against exploring alternatives. By empowering outsiders to engage meaningfully with complex fields, AI tools create new pathways for overcoming these social and psychological barriers to scientific advancement.
The Limitations and Future Potential
Despite its promise, the AI-empowered generalist approach faces significant limitations, particularly the Dunning-Kruger effect. When AI tools provide superficial understanding of complex fields, generalists may develop inflated confidence without appreciating a domain's true depth and nuances.
A physicist's decades of experience provide not just equations but understanding of conceptual subtleties, failed approaches, and boundary conditions. AI can explain quantum concepts convincingly while missing crucial context, leading generalists to propose "revolutionary" ideas that specialists immediately recognize as fundamentally flawed or previously explored.
Combating this requires exceptional intellectual discipline: the confidence to pursue unorthodox questions paired with deep humility about one's limitations. Generalists must constantly question their understanding, seek contrary evidence, and maintain rigorous standards for evaluating their own ideas. This highlights why specialists remain essential. Their expertise provides context for evaluating new ideas, their methodological mastery enables hypothesis testing, and their historical knowledge identifies when "novel" ideas are reinventions of discarded approaches.
The ideal model is collaboration where generalists generate cross-disciplinary insights while specialists refine and validate these ideas. Without verification and peer review, AI-assisted exploration could devolve into an echo chamber of plausible-sounding but flawed concepts. Current AI tools compound these risks with their tendency to produce confident-sounding but incorrect explanations. They cannot (as of yet) independently conduct experiments, gather empirical data, or develop the embodied understanding that comes from direct interaction with the physical world.
Future AI developments could help address these concerns through systems that flag oversimplifications, highlight scientific controversies, or quantify uncertainty. This could help generalists develop a more calibrated sense of their understanding while still benefiting from AI's explanatory capabilities.
Despite these challenges, the potential benefits remain enormous, if approached with appropriate rigor and humility. Scientific revolutions come from rethinking assumptions and crossing disciplinary boundaries, but succeed only when built on solid foundations of understanding.
The next scientific revolution may not come from a lone genius or an elite research institution, but from the combined power of human creativity and artificial intelligence, enabling us to see connections, ask questions, and explore ideas in ways previously impossible.
