Researchers have apparently succeeded in teaching an AI to “think” like humans. They asked an artificial intelligence to hold an internal monologue and thereby made it significantly smarter.

Scientists at Stanford University and researchers at the marketing company Notbad AI have apparently succeeded in teaching an artificial intelligence human-like thinking skills. At least that is the result of a joint study that was uploaded to the document server arXiv.

The research paper has not yet gone through the so-called peer review process. Specifically, this means: The study results have not yet been checked and confirmed by independent scientists. Nevertheless, the studies appear promising.

Artificial intelligence conducts inner monologue

The researchers instructed an AI model to first conduct an internal monologue before answering certain questions. According to the study results, the artificial intelligence then paused to “think”. The AI ​​then ran through various answer options and lines of argument.

The result: According to the researchers, she only spat out the best possible answer at the end of the process. The inner monologue conducted is in turn comparable to the human thought process. Because people usually think first before answering complex questions.

AI is becoming significantly smarter

The researchers used the AI ​​language model Mistral 7B for their investigations. It was trained with seven million parameters and is said to be more powerful than Facebook parent company Meta's Llama language model.

According to the researchers, the AI ​​model performed significantly better after they trained it with the so-called Quiet Star algorithm – also called Quiet Self-Taught Reasoners. According to the study results, after training, the AI ​​was able to increase its answer accuracy from initially 36.2 percent to 47.2 percent.

She was apparently able to almost double her performance on a math test. With an initial accuracy of 5.9 percent, the artificial intelligence achieved a performance of 10.9 percent after the internal monologue. The results still appear to be expandable. However, the researchers are already calling it a success.

Until now, common AI models have not been able to put questions into context and imitate processes. This property otherwise only occurs in the human brain.

Also interesting:


Leave a Reply

Your email address will not be published.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