Trends & Viral Yonhap News Culture · 2026-10-23

AI Building Smarter AI: Tech Pioneers Warn of 'Intelligence Explosion' and Urge Nuclear-Style Oversight

Key Points
  • Over 20 leading AI scientists, including Geoffrey Hinton and Yoshua Bengio, published research analyzing whether AI-driven software development could trigger an 'intelligence explosion.'
  • The study reveals rapid automation gains, such as Anthropic seeing over 80% of approved code written by AI, highlighting growing domestic and global interest in autonomous R&D capabilities.
  • Researchers caution that by mid-2028 AI could automate months-long research projects, urging governments to place independent on-site monitors in AI firms similar to nuclear plant safety inspections.
Summary Can artificial intelligence genuinely build better artificial intelligence, setting off a rapid feedback loop of recursive self-improvement? A landmark paper led by renowned pioneers Geoffrey Hinton and Yoshua Bengio, alongside leading figures from OpenAI, Anthropic, and Microsoft, tackles this urgent question by analyzing the automation of AI research and development and the looming potential for an "intelligence explosion." The report, published under Cambridge University's AI Science and Policy Program (CASP), examines a scenario where AI tools write code, execute experiments, evaluate outcomes, and produce superior next-generation models—which in turn take over future research. Because algorithmic and software refinements can be recycled instantaneously into new cycles without waiting for slow hardware rollouts, progress that once demanded years could theoretically compress into months. Data from Anthropic highlights how quickly this shift is advancing: the proportion of internal code written by AI and approved by human engineers surged from single digits in early 2025 to over 80% by May, while autonomous research and development tasks expanded from 1% to 26% within five months. While researchers emphasize that AI is not yet capable of triggering an uncontrolled intelligence explosion—frequently failing at complex problems solved by experienced human researchers or misreporting task progress—they estimate that by mid-2028, autonomous agents could handle multifaceted research projects that currently take humans months. This acceleration holds immense promise for breakthroughs like accelerated drug discovery, but equally raises catastrophic concerns, ranging from sophisticated cyberattacks and bioweapon proliferation to humans permanently losing governance over automated systems. To preempt these risks before development outpaces oversight, the study urges global governments to mandate transparent reporting from AI labs and install independent, on-site auditors inside major tech firms—drawing direct parallels to regulatory inspectors stationed at nuclear power plants. In tech-driven societies like South Korea, where rapid automation and AI integration are reshaping daily life and industry, this international discourse is sparking deep reflection on digital governance, oversight standards, and the imperative to balance innovation with systemic safety.
Sponsored · Ad