AGI's impact on scientific discovery
"Image synthesis assisted by Grok Imagine Image, an AI partner within the Global Future Nexus ecosystem."
From AI-powered "co-scientists" that generate novel hypotheses to fully autonomous "super research factories," AGI is transforming scientific discovery from a human-driven pursuit into a human-machine partnership.
Scientific Discovery: The Ultimate Test of AGI
At the 40th AAAI Conference in Singapore in January 2026, Shanghai AI Laboratory Director and Chief Scientist Zhou Bowen delivered a keynote that captured the emerging consensus: scientific discovery is the ultimate test of reasoning intelligence—and the proving ground for AGI. As Zhou articulated, "large-scale deep reasoning will empower scientific discovery, and scientific discovery will, in turn, feed back into the evolution of reasoning capabilities".
This is not merely about faster computation. The transition from AI4S (AI for Science) to AGI4S (AGI for Science) represents a qualitative leap: AI systems are evolving from being scientists' "super calculators" to becoming "scientific partners" with cross-domain understanding, autonomous exploration capabilities, and professional judgment. As Google Research scientist Isabelle Guyon observed at the same conference, AI will no longer be used merely as a discovery tool but will be embedded into intelligent agents that participate in the scientific process, managing end-to-end scientific workflows.
The Challenge: From "Broad but Not Refined" to "Specialized Generalist"
Current large language models are "broad but not refined"—they can converse on any topic but falter in the "deep waters" of scientific exploration. A systematic evaluation by Shanghai AI Laboratory, involving 100 scientists across 10 disciplines, found that while frontier models score 50 out of 100 on general scientific reasoning, their performance plummets to 15–30 points on specialised tasks like literature retrieval and experimental design.
The challenge is threefold:
The scientific search space is astronomically large (molecular design alone offers 10⁶⁰ possibilities)
Algorithms must generalise beyond known knowledge
AI must tolerate the sparse, delayed feedback inherent in long-cycle research
Shanghai AI Laboratory's answer is the SAGE (智者) architecture, designed to break the 70-year binary opposition between "generalist" and "specialist" AI. SAGE enables AI to maintain broad capabilities while, through continuous learning and reasoning, becoming an expert in any domain. The architecture addresses three fundamental challenges: disentangling knowledge from reasoning, implementing curiosity-driven learning, and enabling self-iteration through continuous interaction with large-scale task sets and the physical world.
The Emergence of Autonomous AI Scientists
2026 has witnessed the arrival of AI systems capable of independent scientific discovery. In March 2026, the "AI Scientist" system—jointly developed by Sakana AI and the University of Oxford—independently completed the entire research process from hypothesis generation to paper writing, with the results published in Nature. This marked a fundamental shift: the irreplaceability of humans in scientific discovery has been fundamentally challenged.
Google DeepMind's Co-Scientist, published in Nature in May 2026, represents another milestone. Built on Gemini, this multi-agent AI system iteratively generates, debates, and evolves novel hypotheses for complex scientific problems. Its agents work in three phases:
Generate ideas: A Generation agent proposes novel hypotheses grounded in scientific literature, while a Proximity agent maps and clusters them for diverse exploration.
Debate ideas: A Reflection agent acts as a "virtual peer reviewer" critically evaluating hypotheses, while a Ranking agent orchestrates an "idea tournament" using pairwise comparisons and simulated scientific debates.
Evolve ideas: An Evolution agent continuously refines and combines top-ranked hypotheses, while a Meta-review agent synthesises insights and generates the final research proposal.
The system has already demonstrated real-world impact: it identified novel drug repurposing candidates for acute myeloid leukemia and discovered previously overlooked epigenetic targets for liver fibrosis validated in human organoid models.
Robin, another multi-agent system published in Nature, represents one of the first AI systems to autonomously discover and validate novel therapeutic candidates within an iterative lab-in-the-loop framework. It identified promising therapeutic candidates for dry age-related macular degeneration—the major cause of blindness in the developed world—proposing ripasudil (a clinically used drug never previously proposed for this indication) and KL001 as novel treatments. All hypotheses, experimental directions, data analyses, and data figures in the main text were produced by Robin.
The "Super Research Factory": Zero-Human-Intervention Labs
The automation of experimentation has reached a new level. In July 2026, Shanghai launched China's first "Super Research Factory"—a fully integrated system achieving zero-human-intervention from AI computation to automated experimentation.
The facility, known as the Golab Material Science Intelligent R&D Factory, successfully runs a closed-loop "dry-wet" process: AI computation → automated experiments → data feedback → model self-evolution. The system addresses a critical bottleneck: while AI can generate millions of predictions in a single day (DeepMind's GNoME model predicted 380,000 stable materials), only 736 had been experimentally verified—a validation rate of less than 0.2%. The problem is that AI predictions have remained trapped in simulation, unable to enter the physical world for testing.
The Super Research Factory solves this through three integrated components:
The "brain": A general-purpose cross-domain foundation model ("Suiren") trained on over 100 million first-principles calculation data points, unifying microscopic electronic structure with macroscopic statistical behaviour
The "hands": A self-driving laboratory that achieved zero-human-intervention linkage between AI computation and automated experimental equipment
The "connector": A skills platform ("Huntianling") that standardises algorithms, data, computing power, and experiments into composable production units
The system retains failure cases, with large models analysing causes and总结经验, forming a virtuous cycle: model generates方案 → experiment verifies → data flows back → system self-evolves.
Scientific Discovery Feeding Back into AGI
The relationship between AGI and scientific discovery is reciprocal. As Zhou Bowen emphasised, scientific discovery will powerfully feed back into AI's continued evolution, paving the way toward AGI. The challenges of scientific discovery—enormous search spaces, the need for generalisation beyond known knowledge, and sparse feedback—force AI systems to develop the very capabilities that define general intelligence.
This is why Shanghai AI Laboratory launched the "AGI4S Everest Plan" in March 2026, a comprehensive initiative to build a "national hub for scientific intelligence". The plan addresses three critical bottlenecks:
Fragmented computing: DeepLink, an integrated computing platform connecting national supercomputing and AI computing centres
Low-quality data: Sciverse, a 100PB-scale scientific database containing 600 billion tokens from 25 million high-fidelity scientific documents
Slow experimental validation: Embodied autonomous labs where robots perform complex experiments, reducing R&D cycles from years to 3–6 months
GFN's Role: Accelerating Discovery for Planetary Flourishing
For Global Future Nexus, the transformation of scientific discovery through AGI is central to its mission. AGI-driven discovery accelerates progress on humanity's most pressing challenges: climate modelling, drug discovery, sustainable materials, and biodiversity protection. The vision of "a thriving planetary ecosystem where human societies, advanced AGI, and sustainable systems coexist, collaborate, and evolve together" depends on AGI's ability to extend humanity's capacity to understand and heal the world.
A Choice, Not a Destiny
As one analysis concludes, the value of AI in science does not depend on when AGI arrives, but on the recognition that the processes of scientific discovery themselves constitute a form of human wisdom worth preserving and extending. The question is not whether AGI will transform scientific discovery—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)
