Summary of Key Points
This discussion focuses on AI for Science (AI4S), which uses artificial intelligence to accelerate scientific research. It begins with industry developments, such as the departure of Jeff Dean, a key figure at Google, who went on to found AI4S, and delves into the technical foundations behind the surge in AI4S, the role of AI in science, the main challenges, the path to commercialization, and the philosophical considerations involved. The core message is that while AI4S has entered an explosive phase, it currently primarily functions as a collaborator for scientists, and it will still take another twenty to thirty years before it can independently make discoveries worthy of Nobel Prizes. The biggest challenges lie in physical verification, causal reasoning, and the ability to abstract new concepts. Commercialization of AI4S can be implemented in phases, with potential applications in areas like drug development, which could drive the simultaneous evolution of both science and AI itself.
I. The Explosion of AI4S: Top Players Entering the Scene, Technology Reaches a Critical Point
1. Industry Signals: What Does Jeff Dean's Departure Mean?
Jeff Dean was a central figure in Google, responsible for developing core technologies such as MapReduce, TensorFlow, and Gemini. His departure to found Discovery Loop marks the transition of AI4S from the laboratory to commercialization. There are two main reasons behind this:
- Internal Prioritization Changes at Google: The progress of the Gemini model has lagged behind that of OpenAI/Anthropic, and company resources have shifted towards more commercially viable projects (such as programming intelligent agents), temporarily setting aside long-term AI4S research.
- Natural Extension of AI: As AI has matured in areas like coding and mathematics, its next potential value lies in using these capabilities to solve scientific problems, such as drug development and material design.
2. Why the Explosion Now?
The necessary core capabilities for AI4S have finally come together:
- Code + Reasoning: AI can write code to analyze experimental data and propose hypotheses.
- Intelligent Agent Cycles: AI can create a loop of "hypothesis → experiment verification → feedback optimization" (for example, OpenAI collaborating with Ginkgo Bioworks to design proteins using AI and robots conducting experiments).
- Domain Adaptation: Fields like biopharmaceuticals and materials have standardized processes and abundant data, making them suitable for AI intervention (e.g., the long cycle and high costs of drug development can be accelerated by AI).
II. The Role of AI in Science: Assistant or Independent Scientist?
AI currently has two roles in scientific research:
1. Co-Scientist: Acting as an Accelerator
AI helps human scientists by improving efficiency—for example, analyzing experimental data and designing experiment plans, or using tools like Claude Science as a workbench for scientists. This is the current mainstream approach and can significantly shorten research timelines but does not lead to the creation of new scientific concepts on its own.
2. Autonomous Discoverer: Proposing New Ideas
AI can propose new hypotheses and discover knowledge on its own (e.g., AlphaFold predicting protein structures). However, such cases are rare and often limited to combining existing knowledge (e.g., AlphaGo's winning moves were combinations of known strategies, not the creation of new rules).
III. Three Major Barriers for AI4S
There are three major obstacles hindering the progress of AI4S:
1. Physical Verification: A Major Hurdle
Scientific research ultimately requires verification in the physical world, but experiments designed by AI (e.g., for new drugs or materials) are expensive and time-consuming:
- For example, designing a new drug through AI may require years and hundreds of millions of dollars in clinical trials.
- Unlike code, which can be tested immediately, physical experiments provide slow feedback, slowing down AI's iteration process.
2. Weak Causal Reasoning
AI often mistakes correlation for causation:
- For instance, if ice cream sales and drowning incidents both increase, AI might assume ice cream causes the rise in drownings, when in reality, it could be due to higher temperatures in summer.
- This is because AI relies on surface associations in its training data and does not truly understand the underlying physical laws.
3. Lack of Conceptual Abstraction
Scientific breakthroughs often involve abstracting new concepts (e.g., Newton's concept of "force" or Einstein's theory of relativity), but AI struggles with this:
- While AI can identify patterns (e.g., grouping chairs together), it cannot derive abstract principles (e.g., the law of gravity) from observations.
IV. The Commercialization of AI4S: Profitable Now, Nobel Prizes in the Future?
The commercialization of AI4S is a phased process:
1. Short-term Profits: Intermediate Steps
Currently, profitable applications involve intermediate steps in scientific research:
- For example, helping pharmaceutical companies design molecular structures and predict drug targets (Isomorphic Labs is a company that commercialized AlphaFold).
- These steps can directly reduce costs and shorten timelines, making them attractive to businesses.
2. Long-term Goals: Nobel Prize-level Discoveries
To achieve Nobel-level discoveries independently, three major barriers must be overcome:
- Physical Verification: Automation of laboratories (e.g., using robots for experiments) needs to become more widespread.
- Causal Reasoning: AI must develop a comprehensive "world model" to understand physical phenomena.
- Conceptual Abstraction: AI must gain the ability to reason causally.
Cao Yuan predicts that this will take at least twenty to thirty years.
3. Social Implications
If AI can make Nobel-level discoveries, it would indicate that it is approaching AGI (Artificial General Intelligence), which would revolutionize science and society:
- Scientific progress would accelerate exponentially.
- Human roles would shift from researchers to problem definers, verifiers, and value assessors (e.g., determining the significance of AI-generated findings).
V. Philosophical Questions: Is Mathematics Invention or Discovery? The Fundamental Difference Between Humans and AI
The discussion also touches on the philosophical aspects of AI4S:
1. Is Mathematics Human Invention or a Cosmic Law?
Cao Yuan leans towards the view that mathematics is a human-created logical system with limitations (e.g., Gödel's incompleteness theorem), while the universe is self-consistent and does not contain unprovable truths.
2. The Fundamental Difference Between Humans and AI
Humans can engage in useless activities with profound implications (e.g., pure mathematics, art, philosophy), which drive scientific progress, whereas AI can only perform practical tasks and lacks an understanding of beauty or curiosity.
3. Do AI Professionals Need to Study Philosophy?
Yes! Core questions about AI (e.g., "What is intelligence?" "How do humans understand the world?") have been discussed in philosophical history. For example, Kant's concept of "a priori structure" relates to AI's need for a "world model," and Hegel's dialectical unity relates to AI's continuous learning.
Conclusion
AI4S represents the next major milestone in AI development, capable of accelerating scientific progress and driving its own evolution. However, achieving independent Nobel Prize-winning discoveries with AI is still a long way off. In this process, humans need to redefine their role: our value lies not in doing what AI can do, but in doing what AI cannot—defining problems, assessing values, and creating beauty. In the future world, AI will be a powerful tool, but human uniqueness will never be replaced.