Summary of Key Points
This discussion focuses on AI for Science (AI4S), which utilizes AI to assist or conduct scientific research independently, and addresses three critical questions: Why has AI4S suddenly become so prominent? What can it do, and what are its limitations? Where lies the potential for commercialization and the value it offers to humanity? The conclusion is that the surge in AI4S is due to the maturation of AI’s coding capabilities and its ability to streamline research processes. AI is well-suited for solving clear, verifiable scientific problems, but the biggest obstacle is experimental validation, which is costly and time-consuming. Commercialization is also phased: currently, AI can generate revenue from intermediary steps (such as designing drug targets), but it will still take another twenty to thirty years for AI to independently achieve Nobel Prize-level results. The core value of humanity lies in defining research questions, assessing the significance of results, and steering the direction of scientific progress.
1. Why has AI-assisted research only become popular now? – AI has first learned to code and solve mathematical problems
It’s not that science has suddenly become easier; rather, two key aspects of AI have matured:
- Robust coding skills: AI can write code to analyze experimental data without human intervention.
- Process automation: AI can automate the entire research workflow, from hypothesis formulation to experiment design, data analysis, and result adjustment (this is known as “system governance”).
For example, a biological experiment involves three steps: formulating a hypothesis (which requires reasoning), writing code to analyze data (which requires programming), and troubleshooting data anomalies (which requires memory and iteration). Previously, AI lacked one or more of these capabilities; now, it has all of them, creating the perfect conditions for its adoption.
2. Can AI be a helper or a leader in scientific research? – Suitable for clear problems, but not for identifying good research questions
AI is not capable of solving everything; only problems that meet three criteria are suitable for it:
- Computable: The problem can be represented by mathematical models.
- Well-defined: The boundaries of the problem are clear (for example, predicting a protein structure based on an amino acid sequence).
- Quickly verifiable: The outcome can be determined immediately (for example, code can be executed immediately, and mathematical proofs can be verified quickly).
AI can play two roles:
- Collaborative partner: Accelerating experiment design and data analysis (e.g., the Claude Science platform).
- Independent discoverer: Generating new knowledge (e.g., AlphaFold predicting new protein structures).
However, AI has a significant limitation: it cannot identify good research questions on its own. For instance, it cannot make the connection between an apple falling and the law of gravity like Newton did – this requires intuition and abstract reasoning, which AI currently lacks.
3. The biggest bottleneck: Experimental validation is too slow and too expensive
The most challenging aspect of AI4S is experimental validation. Why? While coding and mathematical problems can be verified instantly, scientific experiments require intervention in the physical world, which is time-consuming and costly. Possible solutions include:
- Automated laboratories: Using robots to perform experiments; for example, OpenAI has integrated GPT-5 with a robotic laboratory to automate experiments and provide immediate feedback.
- Improving AI’s accuracy: Making AI’s predictions more accurate so that fewer experiments are needed (e.g., reducing from 10 to 2).
Another hidden bottleneck is the reliability of causal reasoning; AI relies on verbal descriptions to determine cause and effect, which is not as reliable as modeling based on fundamental physical laws. This issue will only be resolved when AI develops a comprehensive “world model.”
4. When can AI4S generate revenue? – Immediate profits from intermediary services, but Nobel Prize-level achievements in 20-30 years
Commercialization is divided into two stages:
- Immediate opportunities: Services such as target discovery and molecular structure design in drug research can be sold, even before the drugs are available on the market.
- Long-term goals: AI will need to independently achieve Nobel Prize-level results, which will take another twenty to thirty years. This is because AI still lacks essential capabilities, such as stable causal reasoning, long-term memory, and the ability to abstract new concepts.
However, AI4S is more than just a application of AI; it can also drive further AI advancements, making it more powerful in other areas such as programming and reasoning.
5. What can humans still do? – Define research questions and assess the value of results
Despite AI’s capabilities, humans play a crucial role:
- Question definition: Humans identify meaningful research topics (e.g., determining which proteins to study for cancer treatment).
- Value assessment: AI’s results may be correct but not meaningful; for example, a computer may prove the Four Color Theorem, but a mathematician may see no insight into it.
- Decision-making: Humans decide whether and how to apply AI’s findings.
Cao Yuan states, “After AI completes its self-evolution, it will elevate humanity to new heights, but we must understand our role: we are the drivers, not just passengers.”
Additional Note
Large companies are investing in AI4S, but its priority is not high compared to other areas such as programming, commercialization, and model expansion. This creates opportunities for startups, as large companies face the “innovator’s dilemma” – once they find a profitable path, they are reluctant to allocate resources to long-term research. Since AI4S is a strategic priority, there is room for innovation by startups.
In summary, AI4S is a slow-moving trend that can generate revenue now, but it will take another twenty to thirty years to truly transform science. Its impact extends beyond science itself, as it will drive further advancements in AI and influence various aspects of human society.
(End of translation)