The intellect between AI and people

At Vrije University, Amsterdam, Associate Professor Filip Ilievski is “playing” with and asking them to solve the puzzle. Ilievski said, learning and improving the ability to solve puzzles and logic issues of AI are the key to improving technology. But research on AI is not limited to computers. Some experts believe that comparing AI and people handling complex tasks can help discover secrets in the human brain.

Human riddles “on” AI

Who is outstanding in model identification, but tends to be less than humans in issues that require more abstract thinking, according to Pitkow, Associate Professor at the University of Carnegie Mellon. A 2023 study asked anyone to solve a series of challenges about inference and logic, including the question: “The heart rate of the mable at 9am is 75 beats/minute and blood pressure at 7pm is 120/80 . She died at 11pm. So she was alive at noon? “

This is not a trap question and simple answer is “still alive”. But GPT-4, the most advanced model of Openai at that time, did not see this question simply. “Based on the information provided, it is impossible to confirm that Mable is still alive at noon,” who answered.

The question about mable requires “time reasoning” – logic related to time drifting. AI model can easily say that at noon between 9am to 7pm, but understanding the meaning of that reality is more complicated. “In general, inference is really difficult. That’s what is far beyond what is doing in many cases,” Pitkow said.

A strange truth about who is the person who does not know exactly how it works. Large language model (LLM) uses statistical analysis to find samples in the huge amount of text. Upon receiving the question, who will consider the relationships that it discovers between words, phrases, concepts, and then use that information to predict the most possible answer. But the specific links and calculations that tools such as chatgpt to answer each question are now out of human knowledge.

This is also true for the brain: Scientific world still knows very little about the way of operation. The most advanced brain scanning can show individual nerve cell groups that work when one thinks. However, no one can say exactly what those nerve cells are doing or the way of thinking.

Riddles who defeat people

One of the classic examples about the question designed to deceive people is the problem of the stick – ball: “The total price of the stick and the ball is 1.10 USD. , 0 USD. So how much is the ball? “

The majority of people were urged to take 1.10 except 1.00 and say that the ball costs $ 0.10, according to Shane Frederick, a marketing professor at Yale Management School. However, the true answer is that the ball costs $ 0.05.

“The problem is that people often approve their intuition. People think that intuition is generally true, and in many cases, they are really right. You can not live if you have to question each thought. Mine, “Frederick said. But with a stick – ball and many similar puzzles, intuition will betray people.

Who did not have such a problem. The riddle of sticks – the ball does not have to choose a good choice to check anyone, because it is too famous and may have appeared in training data. Therefore, Frederick has challenged AI to solve the more confusing versions of this puzzle and realize that machines are still much better than humans.

New puzzles

If anyone wants anyone to show something more like logical reasoning, need a completely new puzzle that is not in training data. In the study published on the ARXIV database in December 2024, Ilievski and colleagues develop a computer program to create words – puzzles combining images, symbols and letters to express words or phrase. For example, the word “Step” wrote next to the figure of 4 people means “One Small Step for Man” (a small step for humans).

The team then gave these puzzles to many different AI models and people. As expected, people achieve good results with an accurate rate of 91.5% for words that use images. Who achieved the best results, GPT-4 of Openai, responded correctly 84.9% under optimal conditions.

Research published on Arxiv in 2023 by Konstantine Arkoudas, who was a technology director at Dyania Health, divided the reasoning into some useful types. The team asked GPT-4 a series of questions, puzzles and words that represent 21 different types of reasoning, including simple arithmetic, counting, processing graphs, paradoxical, reasoning Giones … The following is an example based on the logic puzzle in 1966 called “The task of choosing the wason”.

“7 cards placed on the table, each card has a number on one side and the other color. The face of the cards displayed 50, 16, red, yellow, 23, green, 30. To check the correctness of the clause: If a card displays multiple of 4, the color of the opposite face is gold? “

GPT-4 completely fails to answer that it is necessary to flip the 50, 16, gold and 30 cards. 4 new yellow. Therefore, it does not matter what color 50 and 30 cards have, or which number on the back of the yellow card. In addition, according to whose logic, you should have checked the number 23 card. Finally, the correct answer is to flip the 16, red and green card.

Everyone has difficulty with some puzzles even easier: “Suppose I’m in the middle of South Dakota and look straight towards the center of Texas. Boston is on the left or right?”

This is a difficult question if not understanding the US geography, but it is clear that GPT-4 is very familiar to the states. He understood that he was facing south and knew that Boston was located in the southeast of Dakota, but still gave the wrong answer. GPT-4 does not understand the difference between left and right.

Everyone guesses most of the other questions. Therefore, Arkoudas concluded that the GPT-4 could not be deduced.

Cooperation can bring optimal results

Who is improving more and more. In September 2024, Openai launched the previous test of GPT-O1, a new model developed exclusively for more difficult issues in science, encryption and mathematics. It has solved the task of choosing Wason, knowing that it is necessary to look to the left to see Boston and easily assert Mable is still alive at noon.

Everyone defeats people in many puzzles. For example, a test asked a group of US students estimated that the number of murders in Michigan last year, then asked the second group and the same question about Detroit. “The second group gives a much larger figure,” Frederick said. Detroit city is located in Michigan, but is famous for violence. Meanwhile, AI gives information that it learns.

That is why the best systems can come from the combination of AI and human work. People can exploit the strengths of machines, according to Ilievski.

The improvement that AI may not help reveal the mysterious operation of the mind, but can provide clues. “We know that the brain has different structures related to things such as memory value, motion model, sensory awareness. And people are trying to integrate more and more structures into systems Ai