Artificial Intelligence on the Brink of Humanity’s Last Test
Incredible advancements in artificial intelligence are making it capable of outperforming humans in various intellectual tasks, such as chess and pattern recognition. Experts now suggest we may be just a year away from seeing AI pass what’s being called “Humanity’s Last Test” (HLE). This is a challenge meant for only the sharpest minds of our species, and it may remain unsolvable for AI.
Calvin Zhang, the research director at Scale—an AI firm that developed HLE—expressed admiration for the progress made in building these inference models. He noted that this test will be a significant indicator of how closely AI approaches human intelligence.
Dr. Tung Nguyen, a computer science professor at Texas A&M who contributed a considerable number of questions, emphasized that the HLE serves as a clear measure of the disparity between AI and human cognitive abilities. This benchmark includes 2,500 questions across a plethora of highly specialized fields, ranging from mythology to rocket science, designed to challenge even those with advanced degrees.
More than 1,000 experts in various disciplines contributed to developing the HLE, which is intended to assess knowledge beyond what AI can currently master. According to reports, the aim is to establish an academic standard that only a select few humans might fully grasp.
Interestingly, AI’s performance on the HLE has drastically improved in a short time. For instance, ChatGPT got less than 3% correct on its initial attempt last year, while its competitor, Google Gemini, scored 18.8% after a few months, climbing to over 45% recently.
Zhang is optimistic, believing that AI could soon achieve near-perfect scores. Those hitting close to the 100% mark will be recognized as “all-around experts” within a year.
Kate Olszewska, a product manager from Google DeepMind, shared her thoughts, suggesting that if one is genuinely committed to mastering this challenge, they could reach that expert level fairly quickly.
Such rapid advancements are notable, especially considering the test aims to minimize AI’s advantage. The test developers even offered a $500,000 prize for experts who could create questions difficult to answer through web searches, resulting in over 70,000 submissions.
To ensure that the exam was resistant to AI’s current capabilities, previous questions that could easily be answered by AI were discarded, leaving 2,500 highly challenging inquiries. Contenders might find themselves tasked with translating ancient Palmyrene inscriptions or identifying intricate anatomical structures of birds.
The creators took further precautions by concealing most answers from the AI models to avoid them remembering the responses.
Dr. Nguyen reiterated that while some AI models are showing promising results, the varying performance among them highlights the considerable gap still existing between AI and human intelligence. He pointed out that, while advancements are exciting, intelligence encompasses more than mere pattern recognition; it delves into context, depth, and expertise.
The overarching aim isn’t to defeat AI but to elucidate its strengths and weaknesses. This, hopefully, will pave the way for technology that is safer and more reliable. It also underscores the continued significance of human expertise, especially as AI begins replacing human roles across multiple sectors, including fast food and healthcare.
Nonetheless, AI has demonstrated surprisingly human-like problem-solving capabilities, suggesting that it can go beyond simple memorization. Research from China in 2025 showed parallels between how AI models perceive information and human cognitive processes, particularly in handling language, leading researchers to speculate that AI may be forming human-like conceptual frameworks.
Further exploration revealed a strong alignment between the AI’s understanding and the brain’s activity patterns linked to memory and recognition of scenes.
