구글 딥마인드, 에이전틱 AI, 과학연구 패러다임 자율적 추측 기계로 전환, 검증 병목 현상, 자율 구동 연구실(26.07.17)

00:00 New Agents of Scientific Discovery 01:14 Paradigm Shift in Scientific Discovery 01:45 Multi-Agent Collaborative Architecture 02:38 Virtual Idea Tournament 03:41 Discovery of Drugs to Reverse Liver Fibrosis 06:06 All-round Life Science Achievements 06:19 Physical Verification Bottlenecks 07:48 Disruption of the Academic Ecosystem 09:53 Epistemological Flaws of AI 13:00 Solutions to Overcome Machine Hallucinations 14:28 Autonomous Laboratories (Dark Labs) 15:37 Cost-Aware Cross-Domain Strategies 15:50 Research Workflow Paradigm Shift 16:07 Global Hegemony Competition and National Strategies 17:28 Resolving Verification Bottlenecks and Policy Recommendations The Paradigm Shift in Scientific Discovery and the Rise of the Guessing Machine Historically, scientific progress has been achieved through hypothetical-deductive models, with the intuition and reasoning of human researchers at its center. In the past, artificial intelligence was merely an auxiliary tool to assist humans in isolated areas such as classification or prediction; however, the situation has changed with the remarkable recent advancements in agentic systems based on large-scale language models. AI is now leading scientific innovation as an autonomous speculative machine that coordinates entire scientific workflows, designs complex hypotheses on its own, and learns from experimental failures. The latest frontier models are demonstrating the ability to formulate multi-stage plans and fuse interdisciplinary knowledge, setting out for the true discovery of knowledge. Designing a Multi-Agent Collaborative Architecture A representative example demonstrating how agentic AI evolves scientific hypotheses is Google DeepMind's CoScientist system. This multi-agent system breaks away from linear thinking and features an asynchronous task execution framework in which multiple AI agents assigned specialized roles collaborate and critique one another. The system is modularized into components such as generative agents that search literature to propose hypotheses, reflective agents that critique hypotheses, and ranking agents that prioritize through virtual tournaments, thereby preventing cognitive overload. In particular, it is designed to suppress the hallucination errors inherent to language models by concentrating computational resources on the verification process rather than hypothesis generation. *Empirical Achievements in the Fields of Life Sciences and Medicine* Hypotheses proposed by the guessing machine are proving their validity in real-world physical environments, extending beyond simple text combinations. In a study with a Stanford University research team searching for drugs to reverse liver fibrosis, Coscientist inferred from a vast amount of literature to establish specific epigenetic hypotheses and proposed drugs such as vorinostat. The team's actual experiments showed that while the drugs selected by humans were ineffective, the drugs recommended by Coscientist achieved a remarkable 91% blockage of liver scarring reactions. In addition to this, artificial intelligence is proactively identifying new scientific truths in areas such as identifying combination therapies for the treatment of acute myeloid leukemia and rediscovering the mechanisms of interspecies gene transfer regarding antibiotic resistance. *Intensification of Verification Bottlenecks and Disruption of the Knowledge Ecosystem* The exceptional computational power of guessing machines has, paradoxically, caused a verification bottleneck—a massive physical lag. This is because the speed of infrastructure for physical verification—such as culturing and testing toxicity in actual laboratories—is unable to keep up with the countless hypotheses churned out by AI. Consequently, the academic peer review system is becoming paralyzed. As plausible papers and proposals generated by AI explode, a deadlock in the knowledge flow structure has intensified, preventing truly valuable ideas from being evaluated promptly. Furthermore, as early-career researchers become entirely dependent on AI, there are concerns regarding the risk of a decline in the value of technology, leading to a deterioration of critical thinking skills and scientific judgment across the academic community. *Epistemological Criticism of Complete Autonomy and Overcoming Limitations* Current agentic systems face several epistemological limitations in establishing a fully autonomous scientific discovery framework. Notably, AI is prone to the McNamara fallacy, where it focuses only on superficial problems that are easy to mathematically optimize while avoiding innovative research involving significant uncertainty. Additionally, because it learns only from records of success such as papers and patents, it lacks knowledge of tacit knowledge or failure data from actual laboratory environments, potentially leading to the repetition of meaningless illusions in the face of real-world variables. To o...