In September, Openai presented a new version of chatgpt designed to think through tasks involving mathematics, science and computer programming. Unlike previous versions of the chatbot, this new technology could spend time “thinking” of complex problems before solving an answer.
The company soon said that its new reasoning technology has overperformed the main systems in the sector on a series of tests that follow the progress of artificial intelligence.
Now other companies, such as Google, Anthropic and Deepseek of China, offer similar technologies.
But can Irs really think like a human being? What does it mean to think for a computer? Are these systems really approaching true intelligence?
Here is a guide.
What does it mean when an artificial intelligence system reason?
The reasoning only means that the chatbot spends a little more time working on a problem.
“The reasoning is when the system does an extra job after the question”, said Dan Klein, a professor of computer science at the University of California, Berkeley, and Chief Technology Officer of Scaled Cognition, an artificial intelligence start-up.
It can break a problem on individual steps or try to solve it through attempts and errors.
The original chatgpt immediately answered questions. The new reasoning systems can solve a problem for several seconds – or even minutes – before answering.
Can you be more specific?
In some cases, a reasoning system will perfect its approach to a question, repeatedly trying to improve the method it has chosen. Other times, it can try several ways to face a problem before being satisfied with one of them. Or he could go back and check some work that he did a few second before, just to see if he was correct.
Basically, the system looks for everything that can to answer your question.
This is a bit like an elementary school student who is fighting to find a way to solve a math problem and scribbled several options on a sheet of paper.
What kind of questions do they require an artificial intelligence system?
It can potentially think about anything. But the reasoning is more effective when you ask questions involving mathematics, science and computer programming.
How is a reasoning different from previous chatbots?
You could ask previous chatbots to show you how they had reached a particular answer or to check their job. Since the original chatgpt had learned from the text on the Internet, in which people showed how they had risen to an answer or checked their job, it could also do this type of self -reflection.
But a reasoning system goes further. It can do this type of things without being asked. And it can do it in wider and more complex ways.
Companies call it a reasoning system because it seems that it works more like a person who thinks through a difficult problem.
Why is the reasoning of the AIA important now?
Companies like Opeeni believe that this is the best way to improve their chatbots.
For years, these companies have entrusted a simple concept: more data on the internet pumped in their chatbots, the better they performed those systems.
But in 2024, they used almost all the text on the Internet.
This meant they needed a new way to improve their chatbots. So they started building reasoning systems.
How do you build a reasoning system?
Last year, companies like Openai began to lean heavily on a technique called reforming learning.
Through this process – which can extend over the months – an artificial intelligence system can learn behaviors through extended tests and errors. Working through thousands of math problems, for example, it can learn what methods lead to the right answer and which are not.
The researchers designed complex feedback mechanisms that show the system when he did something right and when he did something wrong.
“It’s a bit like training a dog,” said Jerry Tworek, Openi’s researcher. “If the system goes well, give him a biscuit. If it’s not good, you say” Bad Dog “.”
(The New York Times sued Openi and its partner, Microsoft, in December for violation of the copyright of news related to artificial intelligence systems.)
Does reinforcement learn work?
It works well enough in some areas, such as mathematics, science and computer programming. These are areas where companies can clearly define good behavior and evil. Mathematics problems have definitive answers.
Reinforcement learning does not work as well in areas such as creative writing, philosophy and ethics, in which the distinction between good and bad is more difficult to define. Researchers say that this process can generally improve the performance of an artificial intelligence system, even when he answers questions outside of mathematics and science.
“Learn gradually what reasoning schemes bring him in the right direction and which is not,” said Jared Kaplan, Chief Science Officer of Anthropic.
Are the reinforcement learning and reasoning systems the same thing?
No. Reinforcement learning is the method that companies use to build reasoning systems. It is the training phase that in the end allows the chatbots to reason.
Do these reasoning systems still make mistakes?
Absolutely. Everything that a chatbot is based on probability. He chooses a path that is more similar to the data from which he learned, if such data came from the Internet or that have been generated through the learning of the reinforcement. Sometimes it chooses an option that is wrong or that makes no sense.
Is it a path for a car that corresponds to human intelligence?
Artificial intelligence experts are divided on this question. These methods are still relatively new and researchers are still trying to understand their limits. In the field of the AI, new methods often progress very quickly at the beginning, before slowing down.