Summary of Key Points
This article discusses how AI is transforming scientific research, with the central argument being that AI has evolved from a supplementary tool to a collaborative partner and even an autonomous “AI scientist,” driving a revolution in research paradigms. However, there are also limitations to its capabilities. To fully realize the potential of AI, it is necessary to establish an “AI-native innovation system” that involves a comprehensive reconfiguration in cognitive logic, organizational structure, infrastructure, and evaluation criteria. This system must address three major contradictions currently present in scientific research: organized research versus decentralized exploration, platform-based supply versus personalized needs, and intelligent productivity versus traditional institutional frameworks.
I. What AI Can and Cannot Do in Research
1. What AI Can Do
- Improve Efficiency: AI can process millions of papers worldwide each year to identify cross-disciplinary “hidden knowledge” (e.g., underlying connections between different fields), quickly sorting out research trends, gaps, and new topics. For example, in chemistry, a finely tuned large model found the optimal conditions for a complex reaction after just 15 experiments, saving hundreds of time-consuming trials.
- Solve Complex Problems: AI can tackle problems with an “exponential explosion” of possibilities (such as the Erdős Conjecture in mathematics) or those with uncertainty (where there is no fixed answer, requiring probabilistic reasoning). DeepMind has used AI to solve several mathematical problems with search spaces larger than the number of atoms in the universe.
2. What AI Cannot Do
- Limitations of Computation: For NP-hard problems (for which no optimal solution can be found theoretically), AI can only provide approximate solutions, not a fundamental solution.
- Lack of Theoretical Frameworks: Issues like quantum gravity and the nature of consciousness, without established theoretical frameworks, are beyond the reach of AI based solely on data and computational power.
- Difficulty in Experimental Verification: Problems involving the early universe or the origin of life, for which experimental data is unavailable, pose challenges even for AI.
- Data Scarcity: Rare events or situations with a long-tail distribution (low-frequency occurrences) require sufficient data; without it, AI’s capabilities are limited.
- Logical Limits: Gödel’s Theorem demonstrates that any rule-based system has unsolvable problems, including those that humans can understand but cannot prove using AI.
II. Three Core Conditions for AI to Be Effective in Research
For AI to truly be useful in research, three essential requirements must be met:
1. Bidirectional Encoding: Teams need to convert domain knowledge into a format understandable by AI. For example, AlphaFold’s Evoformer architecture encodes protein structure information into the AI model.
2. Domain-Specific Data and Platforms: Generalized big data is insufficient; specialized “hard assets” are needed. DeepMind used the PDB protein database to train AlphaFold and an autonomous platform that integrates AI design, robotic experiments, and feedback to create a repeatable discovery mechanism.
3. Fast Human-AI Iteration: The process of formulating hypotheses and verifying them must be accelerated. For instance, in chip design teams, engineers and AI experts should collaborate directly to transform physical designs into AI model outputs, breaking down disciplinary barriers.
III. DeepMind’s Success: An Industrial-Grade Research Paradigm
DeepMind’s continuous successes (such as AlphaFold and GraphCast) are attributed to its unique organizational structure:
- Dual Structure:
- Task-Oriented Teams: Small, specialized teams focus on ambitious projects (e.g., predicting protein structures), operating autonomously like independent startups.
- Central Platform: Provides access to tens of thousands of TPU cores and standardized toolkits, eliminating the need for redundant development.
- Industrial-Grade Processes:
- Project Selection: Projects are selected based on three criteria: high difficulty (e.g., akin to landing on the moon), verifiable standards (e.g., protein structure data in PDB), and significant value (e.g., potential to transform a field). Prototypes are tested in simulators before moving forward.
- Execution: Complex problems are transformed into mathematical language that AI can handle. For example, GraphCast uses graph neural networks for weather forecasting, avoiding the computational bottlenecks of traditional partial differential equations.
- Delivery: Projects result in more than just published papers; they also include databases and open-source tools (e.g., the AlphaFold database) for broader use.
IV. Three Major Contradictions Faced by Research
1. Organized Research versus Decentralized Exploration: While AI increases individual researchers’ productivity, it reduces interaction and narrows the scope of knowledge (potentially leading to a “constricted” scientific landscape). A balance between national research initiatives and individual freedom to explore is needed.
2. Platform-Based Supply versus Personalized Needs: DeepMind’s platforms are effective for well-documented problems, but major breakthroughs often arise from unstructured areas that require scientists’ intuition. Therefore, platforms should complement personalized research approaches.
3. Intelligent Productivity versus Traditional Institutions: AI requires interdisciplinary collaboration, but traditional research structures (pyramid-like hierarchies) hinder this. Organizational and evaluation systems must be reformed to accommodate the AI era.
V. Building an AI-Native Innovation System
An AI-native system integrates AI as a core driver, fundamentally transforming scientific research:
1. Cognitive Paradigm Shift: From a linear approach of “hypothesis → experiment → verification” to a data-driven process of “probability generation → prediction.” AI can identify patterns directly from large datasets (e.g., GNoME discovered 2.2 million new crystals).
2. Human-AI Division of Labor: Humans focus on defining problems, assessing value, and ensuring ethical considerations, while AI handles high-dimensional searches and experiment generation.
3. Infrastructure: Establish three foundations: high-quality scientific data (to connect research islands), large-scale computing power, and automated experimental platforms (AI design → robotic experimentation → feedback optimization).
4. Evaluation Criteria: Evaluate not only papers but also AI models, datasets, and toolkits. Both AI and humans should share the benefits of research outcomes (e.g., when AI discovers a new structure, both developers and domain experts should receive recognition).
5. Risk Management:
- If AI generates too many hypotheses quickly for humans to verify, use a tiered screening process (AI preliminary screening → robotic validation → human decision-making).
- To prevent uncontrolled evolution of AI, implement safeguards (e.g., limiting consecutive iterations to three generations and requiring physical verification).
In summary, AI is not just a simple tool; it is a transformative partner that can change the rules of scientific research. To fully leverage its potential, we must fundamentally reframe how research is conducted—both in thinking and organization—to make it more “AI-native.”