Azernews.Az

By Alimat Aliyeva

Researchers have observed a significant rise in artificial intelligence (AI) performance on the so-called “Humanity’s Last Exam.” Over the past year, scores have surged from just a few percent to dozens of percent, AzerNEWS reports.

The “Last Exam of Humanity” was created by specialists at Scale AI in collaboration with the non-profit Center for AI Safety. The exam draws from a massive pool of 70,000 questions prepared by scientists from 50 countries, from which 2,500 were carefully selected. Each question demands expertise roughly equivalent to a doctoral-level understanding. The format emphasizes short, precise answers—making them difficult to find through standard sources.

In 2024, AI models struggled with the exam. ChatGPT scored around 3%, while Gemini and Claude performed slightly better. The results highlighted a stark gap between the capabilities of AI language models and the level of human scientific specialists.

Today, the landscape has shifted dramatically. According to the developers, Gemini now scores 45.9% and Claude 34.2%. Calvin Zhang, a representative of Scale AI, suggested that some models could reach 100% within months or a year. When that happens, he says, it will be necessary to design questions for which even human experts do not have ready answers.

Despite these advances, the developers stress that AI will not completely replace humans. Certain tasks—especially those requiring unconventional solutions, creativity, or work in highly specialized scientific fields—still rely on human intuition and insight.

Interestingly, some experts speculate that once AI reaches near-perfect scores, the exam itself could evolve into a tool to explore new frontiers of human knowledge, challenging both machines and researchers to push beyond what is currently known.