Summary of Key Points
The central argument of this article is that current AI is adept at making precise predictions within the existing scientific framework (like a map with all details filled in but unable to provide direction), yet it lacks the capability to drive “paradigm shifts” – those groundbreaking discoveries that fundamentally rewrite the rules of science (such as the theory of relativity or evolution). The article warns that over-reliance on such AI can lead to the risk of “extraordinary science”: we become faster at working within existing models but lose the ability to pose entirely new questions. To change this, we need to develop “visionary AI” that can think beyond these frameworks; moreover, we must use AI to study the underlying principles of science itself (meta-science) in order to create conditions conducive to paradigm shifts.
1. The Map Metaphor: Why AI Cannot Achieve Major Innovations
The article uses two map examples to illustrate the limitations of AI:
- Borges’ Perfect Map: A map as large as an empire, with 100% accuracy in details, but completely useless for navigation. Current AI is similar in that it is trained on vast amounts of data (such as trillions of texts or protein structures) and can make precise predictions within existing frameworks (for example, AlphaFold predicting protein folding), but it cannot “redraw the map”.
- The Revolution of the London Subway Map: Before 1933, subway maps were geographically accurate, with central stations crowded together and suburbs occupying large areas, making them difficult to use. Harry Beck abandoned geographical accuracy and created a map with colored lines and evenly spaced stations (like a circuit diagram), which became much clearer and more useful. This represents a paradigm shift – but AI cannot do what Beck did; it does not voluntarily abandon old frameworks to create new, more useful structures.
2. “Extraordinary Science”: The Hidden Trap of AI
“Extraordinary science” refers to the situation where our predictive abilities within existing models increase, yet our ability to pose new questions declines. The article provides several examples:
- The Cholera Epidemic: In 1849, during the cholera outbreak in London, Dr. Farr mapped the areas with higher mortality rates and attributed it to “miasma” (harmful air). If AI were used on Farr’s data, it could predict the next outbreak locations precisely, but it would never identify the real cause – contaminated river water (later solved by Pasteur’s theory of bacteria).
- DeepMind’s GNoME: It identified 2.2 million new materials, most of which were just substitutions of known elements (e.g., using elements adjacent in the periodic table), without breaking out of old frameworks.
- The Limits of AI-Assisted Research: Studies show that scientists using AI publish more papers and get more citations, but the range of their research topics has narrowed by 5%. AI tends to guide them towards existing fields with abundant data rather than exploring new directions. The risk is that we mistake “more details” for a deeper understanding, thus missing out on major ideas that could change the world.
3. The Secrets of Paradigm Shifts: How Humans Do It
The article summarizes three key factors behind paradigm shifts using examples from Einstein and Darwin:
- Prioritizing Simplicity: Einstein replaced the complex theory of “ether” with two principles (physical laws being the same in all reference frames and the constancy of the speed of light); Maxwell unified electricity and magnetism with four equations. Good paradigms are often simple and can explain more phenomena.
- Interdisciplinary Analogy: Darwin was inspired by Malthus’s economics (competition for resources) to propose the theory of evolution; Feynman realized that heat flow, fluid dynamics, and electrostatics share the same set of equations. Cross-disciplinary analogies can provide new perspectives.
- The Courage to Step Out of Frameworks: Einstein, an outsider to the academic community, was not constrained by traditional notions of “ether”; mathematical genius Poincaré, on the other hand, failed to develop the theory of relativity even though he calculated the same mathematical results.
- Tolerance for Imperfection: Darwin’s theory of natural selection initially lacked a mechanism for heredity (he proposed the incorrect “pangeneratism”), but its core idea was strong enough to be refined by the later discovery of genetics.
4. How to Build Visionary AI
To enable AI to drive paradigm shifts, the article suggests the following approaches:
- Pursue Simplicity: The AI Feynman system used “symbolic regression” to find the simplest equations that explain data, successfully rediscovering 100 equations from Feynman’s physics lectures (29 more than previous software).
- Find Interdisciplinary Analogies: AI can identify structural similarities across different fields (similar to how Darwin connected economics and biology).
- Combine Physical Experiences: Multimodal AI (handling vision, touch, movement) or autonomous laboratories (AI + robots) can combine abstract ideas with real-world feedback (like Einstein’s thought experiments of riding on light).
- Human-AI Collaboration: Humans have the breadth of multiple senses (seeing, hearing, touching, thinking), while AI has the depth to process large amounts of data. Together, they can uncover analogies that AI alone cannot find.
5. Meta-Science: Using AI to Study Science Itself
For AI to drive paradigm shifts, we must first understand how science works. This is what “meta-science” entails:
- AI-Simulated Scientific Environments: We cannot conduct controlled experiments in real laboratories (e.g., comparing whether small or large teams are more innovative), but AI can simulate different research environments (team size, hierarchical structures, communication patterns) to determine which conditions foster paradigm shifts.
- Insights from Historical Experience: Institutions like Bell Labs and Xerox PARC produced groundbreaking results because they were small teams with the freedom to pursue seemingly useless research. AI can simulate these conditions to verify their effectiveness.
- Key Conclusion: AI will not automatically bring about paradigm shifts; we need to design environments conducive to innovation (e.g., through institutional incentives and team structures), and AI can help us identify these conditions.
Final Thoughts
The article is not against AI but hopes that it will evolve from a predictive tool to a partner in innovation. To achieve this, we must not only improve AI technology but also use it to study the fundamental principles of science. After all, if we want AI to help us draw new maps, we first need to know what kind of maps are useful.