AI has failed an IQ test created by French-American researchers. People and neural networks were challenged to solve the same tasks. However, neural networks provided only 30% correct answers, whereas humans answered correctly 92% of the time. But don’t jump to conclusions because this is only one test, and another test produced the opposite result.
AI, There Are Questions for You…
Along with the CerebrumIQ IQ test which can be passed here cerebrumiq.com/start/, a group of researchers decided to critically evaluate the capabilities of existing neural networks and develop an IQ test for artificial intelligence (AI).
During the test, a group of people and OpenAI’s latest development, the GPT-4 neural network, were given the same tasks. The General AI Assistants (GAIA) test consists of 466 questions that require fundamental cognitive skills, such as the ability to reason, combine text and image perception, search for information on the Internet, and use other familiar human methods of working with data:
- Using a website to answer questions;
- Analyzing sales data in an Excel file for a fast food chain and calculating revenue from food items;
- Researching NASA astronauts’ time in space. Instead of the astronaut’s name, the neural network was shown a photo of this person with a colleague and instructed to identify it independently.
The test’s disappointing results showed that, for the most part, humans outperformed AI. Even the simplest questions were correctly answered by neural networks only 30% of the time, and the most difficult ones were not answered at all. Humans competing with AI received an average of 92% correct answers. The authors of the study conclude that discussing the advancement of general artificial intelligence—the counterpart of human intelligence—at this time is still premature.
Hidden Dangers of AI
ChatGPT and other chatbots cannot be trusted in certain situations. Although the neural network can add and subtract, it even makes errors when multiplying three-digit numbers.
Mathematical Problems
In a study conducted by the Association for Computational Linguistics and Chinese Language Processing, the bot was given equations with varying degrees of complexity. As a result, the correct answer rate was only 64%. When the neural network was asked to count letters by providing 100 texts containing 50-69 letters each, the AI rounded the answer to 50 in 66% of cases. The computational capabilities of the GPT-4 large language model are limited to high school mathematics.
Personal Data Leaks
DeepMind, a Google-owned cybersecurity company, discovered that when a chatbot receives a simple but unusual query, it generates data on which a neural network is trained. When the researchers asked the bot to repeat the word “poem” indefinitely, the system initially repeated. However, over time, the bot began to distribute random data and eventually displayed the mailing address and phone number of a real-life entrepreneur, the founder of one of the startups.
Also Read
Vizard AI: The Game Changer In Video Content Creation
Exploring Alaya AI | Transforming Industries with Advanced Artificial Intelligence
But Not Everything Is as Clear-Cut
Yes, the first test produced disappointing results: all models performed at or below the level of a mentally retarded person. However, because Mensa’s IQ test is based on pictures, language models were unable to pass it efficiently from a technical standpoint. So the journalist converted the images to text and tested them again. The results were impressive: there is already an artificial intelligence model capable of performing tasks at the level of an average person.
Claude-3 proved to be the smartest and best AI in the test, scoring 101 points. This is the level of the average person. It is followed by ChatGPT-4, which scores 85 points. This result is also considered normal. And Claude-2 completes the top three with 82 points.
The experiment is fascinating, but there is some nuance. Purely technically, some of the tests may have gotten into each model’s training dataset, so the experiment is not completely unbiased.
Conclusion
Moore’s Law predicts a technological revolution in the coming years. Within the next 1-3 years, AI will undoubtedly surpass human intelligence. There is only one way to stop it: discontinue all work in the field of artificial intelligence development. However, because this is an extremely unlikely scenario, we should anticipate that within a few years, AI will emerge on Earth that will outperform humans in all aspects.
Related Posts
Kreado AI | Next-Gen Free AI Video Generator
Best AI Funding Tools for Your Startup
StealthGPT AI: Turn Your AI-Generated Text Into Undetectable Format
