Stanford University’s AI researchers have discovered that current tests are not adequately measuring the skills of artificial intelligence. The HAI Research Institute at Stanford releases annual reports to investigate different versions of AI and their capabilities. Their latest report, released in April 2021, shows that the best AI can now compete with humans in tasks such as image classification, understanding English text, and deducing everyday scenarios based on visual input.
Over the past decade, AI has made significant progress in its capabilities, particularly in math problem-solving. Stanford’s MATH test, which consists of 12,500 math problems, demonstrates this improvement. In a recent test, an AI solved 84.3% of the problems correctly. Despite these advancements, AI still faces challenges such as “hallucinations” in large language models where it can make errors or misinterpret information. In areas like demanding mathematics and design, AI still needs to improve.
To better measure AI abilities and identify areas where humans still outperform them, Stanford University is currently working on developing new tests. They also anticipate the introduction of GPT-5 language model in the near future that may influence future tests and advancements in AI capabilities.
