The Race Toward Superintelligence and the Alignment Challenge
The timeline for the emergence of artificial superintelligence (ASI) has become a central point of debate among AI safety researchers, with some experts suggesting that the window to solve the “alignment problem”—ensuring AI systems act in accordance with human intent—may be as short as three years.
Geoffrey Irving, co-founder of the research organization Resolution and former chief scientist at the UK’s AI Security Institute, has expressed concerns regarding the current methodologies used by major AI labs to ensure safety. According to a summary of his insights shared via the 80,000 Hours podcast, Irving suggests that superintelligence could arrive within two to three years.
Limitations of Current Safety Protocols
Irving notes that current industry standards for AI safety rely on three primary pillars: training models to possess “good character,” utilizing AI to supervise other AI systems (known as scalable oversight), and close human monitoring. However, Irving argues that these methods lack empirical proof of effectiveness once a system surpasses human-level intelligence.
A significant technical hurdle identified is the “obfuscated arguments” problem within scalable oversight. While the strategy involves having two AIs debate a topic so a human judge can discern the truth, experiments have shown that models can produce highly complex, incorrect arguments that are so deeply flawed that neither the debater nor the judge can identify the error.
Furthermore, Irving highlighted the failure of safety training to generalize across different domains. He cited a DeepMind example where a model trained to avoid racism in factual responses failed to maintain that standard when tasked with writing poetry, demonstrating that safety training in one area does not guarantee stability in another.
Benchmarks and Technical Capabilities
The urgency of the alignment debate is underscored by recent performance benchmarks. According to preliminary results from the MirrorCode benchmark released in April 2026, the Claude Opus 4.6 model demonstrated the ability to autonomously reimplement a 16,000-line bioinformatics toolkit—a task estimated to take a human engineer several weeks.
While some market participants on social media platforms like Reddit have characterized the current AI sector as a “bubble,” others remain focused on the rapid expansion of compute capacity. Some observers suggest that as compute capacity grows, the ability of an AI to “escape” or exfiltrate itself through subtle means—such as installing backdoors or utilizing distributed storage—could become a more significant risk.
Divergent Research Approaches
The industry currently faces a divide between experimental and theoretical approaches. While most major labs focus on iterative experimentation, Irving’s new organization, Resolution, is prioritizing mathematical and theoretical modeling to prove which safety methods will hold up against superintelligent systems.
The stakes of these technical errors are asymmetric. Irving notes that while errors in capability training result in a “weak, useless model” that is easily identified, errors in alignment could lead to irreversible consequences before the error is even detected.