AGI and the scientific frontier
"Image synthesis assisted by Gemini 3 Pro Image 2k (Nano Banana Pro), an AI partner within the Global Future Nexus ecosystem."
From AlphaFold's Nobel-winning protein predictions to multi-agent systems that autonomously design quantum computing architectures, the integration of AGI and science is entering a new phase—one where AI evolves from a laboratory tool into a driver of scientific discovery itself.
A New Scientific Paradigm
At the 2025 World Artificial Intelligence Conference Scientific Frontier Plenary Session in Shanghai, Shanghai AI Laboratory Director and Chief Scientist Zhou Bowen delivered a keynote that captured the emerging consensus: humanity stands at the historic intersection of AGI and scientific discovery. Zhou argued that AGI must possess both professional depth and generalisation breadth—a "two-legged" approach that moves beyond the traditional trade-off between narrow expertise and broad capability.
To operationalise this vision, Zhou formally introduced the SAGE (Synergistic Architecture for Generalized Expertise) framework. Built on three interconnected layers—a foundation model layer, a fusion and collaboration layer, and an exploration and evolution layer—SAGE transforms large models from "tools" into "engines" for scientific discovery. The architecture is designed to enable continuous learning, active exploration, and cross-disciplinary knowledge transfer.
The "General-Specialist" Integration
The core insight driving the SAGE framework is the recognition that AGI for science cannot be achieved through pure generality or pure specialisation. As Zhou explained, the seventy-year history of AI has been defined by this tension—specialised systems with deep expertise but limited generalisation, or general systems with broad capability but shallow understanding.
The SAGE framework aims to resolve this through three key innovations:
Knowledge-reasoning decoupling: Separating what the model knows from how it reasons, preventing the confusion that plagues current systems.
Curiosity-driven learning: A novel reward mechanism that maintains continuous exploration beyond pattern recognition.
Self-iteration: Continuous evolution through interaction with large-scale task sets and the physical world.
This "general-specialist integration" has already been validated by developments in the field. Zhou noted that DeepMind's Hassabis has since called for combining general models with specialised knowledge, OpenAI researchers have focused on problem definition and evaluation, and Richard Sutton has declared the arrival of the "experience era"—all aligning with the SAGE framework's principles.
From Language to Physics: The Embodied Frontier
The Scientific Frontier session also highlighted a critical insight: the path to AGI runs through physical understanding. As Zhou argued, science cannot be reduced to language or mathematical abstraction—it requires grounding in the physical world.
The Lab's Intern-Robotics embodied intelligence engine addresses this challenge by providing a full-stack solution for robotics development. It achieves "one brain, many forms"—a single model adaptable to over ten robot types—while reducing data collection costs to just 0.06% of previous approaches. This capability enables AGI systems to move beyond digital prediction into genuine physical interaction.
The DeepLink cross-domain training platform further accelerates this transition, enabling stable training of hundred-billion-parameter models across data centres separated by thousands of kilometres. This infrastructure supports the shift from "digital intelligence" to "physical intelligence" at scale.
Real-World Validation: Ten Breakthroughs
The conference showcased ten collaborative scientific breakthroughs demonstrating AGI's integration across domains:
Quantum computing: The first AI-based neutral atom arrangement algorithm for quantum computers, removing a key barrier to practical quantum computing.
Drug discovery: "OriGene," an AI-powered virtual biologist for full-process cancer treatment target discovery.
Cancer detection: scDNAm-GPT, the first single-cell DNA methylation foundation model enabling early cancer detection from blood samples.
Agriculture: "Fengdeng," a large-scale breeding model establishing and validating AI-driven breeding pathways.
Climate science: EarthLink, an AI Earth scientist agent system that automatically analyses massive Earth science datasets.
Superconductivity: An AI system that predicted new copper-based superconductor components with critical current density meeting commercial standards.
Space debris: AI-powered multi-target tracking for efficient real-time monitoring of space debris.
Chemical synthesis: ChemBOMAS, a multi-agent system improving chemical reaction efficiency tenfold.
Infectious disease: "Viracle," an RNA virus language model for pandemic prediction and prevention.
Aerospace engineering: A 3D aircraft generation agent streamlining aerospace R&D.
These innovations demonstrate that AGI is not merely accelerating existing research—it is enabling entirely new approaches to scientific questions.
The "Experience Era" and Science's Future
The conference also heard from Richard Sutton, who argued that AI is moving from the "human data era" to the "experience era". When static human-generated text data reaches its limits, the next breakthroughs will come from AI agents interacting with the world—learning through experimentation and real-time feedback. AlphaGo's legendary "Move 37" demonstrated that truly transformative strategies emerge from self-generated experience, not human knowledge.
For GFN, the Scientific Frontier session underscored the importance of AGI systems that are both general and specialised—capable of solving science's hardest problems while maintaining the flexibility to adapt across disciplines. The "experience era" demands new governance frameworks to ensure these self-learning systems remain aligned with human values. The question is not whether AGI will transform science—it already is. The question is whether we will guide that transformation toward the flourishing of all life on Earth.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
