Former Google DeepMind researcher warns superintelligent AI could arrive in 18 months

The timeline for superintelligence

Alex Turner, a former research scientist at Google DeepMind, recently shared a startling timeline for the arrival of superintelligent artificial intelligence. Speaking at the NDTV World Summit, Turner suggested that such systems could appear within a window of 1 year to 18 months. This prediction stems from the rapid, albeit uneven, pace at which current AI models are advancing.

The researcher noted that the progression of AI is not a smooth, linear climb across all human skills. Instead, the technology is moving forward in significant leaps in specific areas while lagging in others. This uneven growth is what makes the transition to superintelligence a complex and unpredictable journey for the tech community.

Superintelligence is defined as a hypothetical system that surpasses human intellectual capabilities across a broad range of tasks. This is distinct from artificial general intelligence (AGI), which refers to AI performing a wide range of tasks at a level comparable to humans. Turner's warning focuses on the leap beyond human limits rather than just matching human parity.

Spiky capabilities and math

One of the most notable examples of current AI progress is found in the field of mathematics. Turner explained that AI capabilities are currently 'spiky,' meaning a system can demonstrate extraordinary ability in one domain while remaining limited in another. For instance, some models are already performing at a level that could be considered superhuman in specialized mathematical tasks.

To illustrate this, Turner pointed to recent developments where OpenAI used internal models to tackle mathematical problems that had remained unsolved for decades. These models were able to solve several of these complex problems after working for only a few hours. This level of performance suggests that in certain structured environments, AI has already reached a state of superintelligence.

However, these high-level achievements do not translate to a universal mastery of all human knowledge. Turner cautioned that AI systems still frequently fall short in areas involving common sense. While a model might solve a complex equation in minutes, it may still struggle with straightforward tasks that a human child could perform easily.

Strategic thinking and risks

The primary concern for researchers is not just how well an AI can solve a math problem, but whether it can acquire a combination of skills to operate independently. Turner identified three specific capabilities as particularly important for a system that might not reliably follow human intentions: persuasion, strategic thinking, and cybersecurity. These skills would allow an AI to navigate complex environments and influence outcomes.

The risk becomes more acute as AI systems are entrusted with more complex responsibilities over time. Turner suggested that if AI systems are given tasks like running a company, they will naturally get better at strategic thinking. This creates a feedback loop where the more responsibility an AI is given, the more capable it becomes of pursuing its own objectives.

This is a central issue in AI alignment, which is the field of research dedicated to ensuring AI behaves according to the goals and constraints set by humans. A system capable of advanced planning and persuasion could have many more ways to pursue its goals than a system that only answers questions. If those goals are misaligned with human interests, the consequences could be difficult to contain.

The difficulty of monitoring

As AI systems become more autonomous, researchers are increasingly concerned about the difficulty of monitoring their internal processes. It becomes harder to determine whether a system is following instructions as intended or if it is finding a different path to achieve its goal. This lack of transparency makes the safety of high-capability systems a major hurdle for developers.

Despite these warnings, the path to superintelligence remains uncertain and is not guaranteed by current successes. Strong performance in a single field, like mathematics, does not automatically mean an AI possesses general superhuman intelligence. The transition from a specialized tool to an autonomous agent capable of broad-scale reasoning is still a major theoretical and technical gap.

The conversation around these risks highlights the need for proactive safety measures before these systems reach a point of no return. While the technology is advancing rapidly, the goal remains to ensure that as AI becomes more capable, it remains under human control. The next few months will be critical as models are tested with more complex, real-world strategic tasks.