Anthropic CEO wants to open the black box of AI models by 2027

Anthropier CEO Dario Ameudi published an article on Thursday highlighting the little understanding of the internal work of the world’s artificial intelligence models. To address this, Amodei set an ambitious goal for Athrubor to discover most of the problems of the artificial intelligence model by 2027.

Amodei admits the next challenge. In the “urgency of the interpretation”, the CEO of Anthropor says that Antarbur has made early breakthroughs to track how the models reach their answers – but it confirms that more research is needed to decode these systems with their growth more powerful.

“I am very concerned about spreading such systems without dealing better on the ability to interpret,” Amodei wrote in the article. “These systems will be completely essential to the economy, technology and national security, and will be able to a lot of self -rule that I consider it unacceptable mainly that humanity is completely ignorant of how it works.”

Anthropor is one of the leading companies in mechanical interpretation, a field that aims to open the black square for artificial intelligence models and understand the reason for making the decisions they make. Despite the rapid performance improvements in artificial intelligence models in the technology industry, we still have a relatively little idea how these systems reach decisions.

For example, Openai has recently launched new AI Relativeing ​​models, O3 and O4-MINI, which work better in some tasks, but also hallucinations more than their other models. The company does not know the reason for its occurrence.

“When the Importer IQ system does something, such as summarizing a financial document, we have no idea, at a specific or accurate level, and why he takes the options he makes – why he chooses certain words on others, or why he makes a mistake sometimes even though it is usually accurate.”

In the article, Amodei notes that the co -founder of the anthropologist Chris first says that artificial intelligence models “are cultivated more than they were built.” In other words, artificial intelligence researchers have found ways to improve the intelligence of the artificial intelligence model, but they do not know the reason completely.

In the article, Amodei says it may be dangerous to reach AGI – or as it calls it, “a country of geniuses in the data center” – without understanding how these models work. In a previous article, Amodei claimed that the technology industry can reach such a landmark by 2026 or 2027, but he believes that we are in a state of firm understanding of these artificial intelligence models completely.

In the long run, Amodei says that the anthropor is, mainly, to perform “brain tests” or “magnetic resonance imaging” for modern AI models. These tests help determine a wide range of problems in artificial intelligence models, including their tendencies to lying or searching for power, or weakness of any weakness, he says. He added that this may take from five to 10 years to achieve it, but these measures will be necessary to test and publish artificial intelligence models in Antarbur.

Anthropor has made some research breakthroughs that allowed her to better understand how artificial intelligence models do. For example, the company recently found ways to track the paths of thinking about the artificial intelligence model, what the company calls, and circles. Anthropor has identified a single circle that helps artificial intelligence models to understand the American cities where the United States is located. The company found only a few of these circles, but the estimates there are millions of artificial intelligence models.

Antarubor was investing in the same interpretation research, and it recently invested its first investment in an explanation company. Although the interpretation is largely seen as a field of safety research today, Amodei notes that in the end, it explains how artificial intelligence models reach their answers as a commercial advantage.

In the article, Amodei Openai and Google DeepMind called to increase its research efforts in this field. Besides the friendly batch, the CEO of Anthropic asked governments to impose “lighting light” regulations to encourage interpretation research, such as companies’ requirements to detect safety and security practices. In the article, Amodei also says that the United States must put the chips export controls to China, in order to reduce the possibility of a race of artificial intelligence outside of control.

Antarbur has always emerged from Openai and Google for safety. While other technology companies pushed the controversial safety bill in California, SB 1047, Anthropor issued modest support and recommendations for the draft law, which would have set the safety reporting criteria for AI Frontier developers.

In this case, Antarbur appears to be pushing for an effort at the level of industry to better understand artificial intelligence models, not only to increase their capabilities.