AI can surpass human intelligence in some tasks: here’s what they are

AI can surpass human intelligence in some tasks: here's what they are

Since 2017, every year, experts from Stanford University have published a detailed analysis on the evolution of artificial intelligences and on the increasingly central role they are carving out for themselves in the lives of each of us. In the AI ​​Index Report 2024, which has just appeared on the academic center’s website, we read that AI is now truly capable of surpass human intelligence. Albeit in very specific areas.

The report produced at Stanford is extremely detailed, so much so that it runs across 502 pages, addressing all the segments in which AI is actually used. The report also talks about responsibility of AI and the consequences that arise from its use as well as the performance of the various models.

AI can outperform humans in some tasks, but not all

Experts from Stanford University first compared the scores obtained by humans and AI in 9 types of tests. The objective of these tests was to measure the performance in language processing and image processing.

The newly published study highlights that AI succeeds higher scores to humans on tasks such as simple language management and image recognition, while humans retain higher scores on tests of visual common sense reasoning (VCR) and advanced mathematics skills (MATH).

VCR stands for “Visual Common Sense Reasoning“: this is a type of test used to evaluate the capacity of a AI system in understanding and solving common-sense reasoning problems through the analysis of images or visual scenes.

These tests are designed to evaluate whether an AI can actually understand common situations, identify relationships between objects or events in images, and draw logical conclusions based on visual information. For example, you could ask the AI ​​to solve problems like “What is the missing object in this scene?” o “What is the logical sequence of events that will occur based on the image provided?”.

The “advanced level mathematical skills” (MATH) refer to a set of mathematical skills and knowledge that goes beyond basic concepts and covers more complex and advanced topics. These skills typically include undergraduate or graduate-level mathematical concepts; for example, advanced calculus, advanced algebra, mathematical analysis, advanced geometry, probability, and advanced statistics.

Companies are taking the lead over academic institutions in developing cutting-edge AI

Another noteworthy aspect that emerges strongly from the Stanford University study is that well 51 models for machine learningamong the most relevant of 2023, were designed, developed and created by private companies, 21 involved collaboration between companies and academic institutions, only 15 were developed by academic institutions.

Overall, US entities have developed 61 models; followed by China with 15, France with 8 and Germany with 5.

The costs of training generative models are becoming extremely high

We have written it in many of our articles and the CEO of ARM, Rene Has, also recently reiterated it: artificial intelligence consumes a lot of energy.

The computing power (measured in FLOPs, Floating Point Operations per Second; represents a measure of the computing power of a system, in particular the ability to perform floating-point arithmetic operations in one second (such as additions, subtractions, multiplications and divisions) of the machines used to train machine learning models is constantly increasing . Generative models developed by companies are generally more powerful than academic ones.

As a result, i training costs of machine learning models are increasing as the complexity of the models advances. Suffice it to say that only training GPT-4 (OpenAI) cost something like $78 million; Google spent $191 million training Gemini Ultra.

On the other hand, everything must be commensurate with theincrease in productivity that AI solutions can ensure. For more than half of companies, in any sector, theAI has reduced costs and increased turnover.

There is a lack of standardized assessment criteria for AI safety

Stanford also puts in black and white that AI can represent a threat. The number of incidents involving AI in some way is increasing, and Stanford University supports the need to “standardized evaluation criteria” to properly weigh AI safety.

Cases of fraud carried out using AI or the spread of deepfakes using intelligent image generation systems are already regularly recorded.

AI is accelerating scientific progress

In recent years, AI has proven to be much more than just a technological innovation. Rather, it proved to be a powerful catalyst for the scientific progress in multiple sectors. With increasingly sophisticated algorithms and access to enormous amounts of data, AI is revolutionizing scientific research in ways that were unthinkable just some time ago.

One of the main areas where AI is influencing scientific progress is in its ability to to analyze e interpret large amounts of data quickly. With machine learning algorithms and deep learning techniques, AI can spot patterns and relationships in data that would be difficult or impossible for humans to grasp. This is especially useful in fields like biology, where analysis is large dataset it can lead to new discoveries in the field of personalized medicine or biotechnology. AI is also demonstrating its full potential in discovery of new materialsin chemistry, in medical research and in many other sectors.

The number of people who believe they have a good understanding of AI increased from 64% in 2022 to 67% in 2023. The number of people who believe that AI-based services will change their lives in the next 3-5 years is increased from 60% in 2022 to 66% in 2023.

Opening image credit: – Digital43

Leave a Reply

Your email address will not be published. Required fields are marked *