Anthropic CEO DARIO AMODEI has published an essay that emphasizes how much researchers understand the internal work of the world’s best AI model on Thursday. To solve this, Amodei has set an ambitious goal to stably detect most AI models by 2027.
AMODEI recognizes the challenge in the future. In the “emergency of interpretation,” CEO Manchropic has done a breakthrough in the early days of tracking how the model reaches the answer, but emphasizes that as these systems grow more powerful, they need much more research to decipher this system.
“I’m very concerned about distributing such a system without dealing with the possibility of interpretation,” Amodei said. “This system will be the center of economy, technology and national security, and I don’t think that humanity is completely ignorant about how humanity works because of too much autonomy.”
Anthropic is one of the pioneering companies in the mechanical interpretation, which is a field that opens the black box of the AI model and understands why they make decisions. Despite the rapid improvement of the AI model of the technology industry, we still rarely know how this system reaches the decision.
For example, Openai has recently launched O3 and O4-Mini, a new reasoning AI model that is better performed in some tasks, but has caused more hallucinations than other models. The company doesn’t know why that’s happening.
“I don’t know why I make a mistake,” he said, “When the creation AI system is doing the same thing as summarizing financial documents, why you choose at a particular or accurate level, why you choose a specific word than others, and why it’s usually accurate.
AMODEI said that AI Models is growing more than the AI model. In other words, AI researchers found a way to improve the AI model intelligence, but they don’t know why.
In this article, AMODEI says that it may be dangerous to reach AGI without understanding the way of operation of these models or to be called a “genius state of the data center.” In the previous essay, AMODEI insisted that the technology industry could reach such a milestone by 2026 or 2027, but we believe that it is much more by fully understanding these AI models.
AMODEI says that in the long run, Anthropic is essentially intended to perform “brain scan” or “MRI” of the cutting -edge AI model. This examination will help to identify a wide range of problems, including lies, power or other weaknesses in the AI model, he said. This may take five to ten years to achieve, but this action will be necessary to test and distribute Anthropic’s future AI models.
Anthropic has created some research innovations that can better understand how the AI model works. For example, the company recently found a way of thinking about the accident route of the AI model and a way to track the circuit. Anthropic, which helps the AI model to understand the US cities in the United States, identified a circuit. The company has found some of these circuits, but estimates that there are millions of people in the AI model.
Anthropic has invested in analytical research itself and has recently made its first investment in interpretation. In the essay, AMODEI asked for increasing research in the field through Openai and Google Deepmind.
AMODEI requires that the government will impose the “Light-Touch” regulations to encourage the study of analysis, such as the requirements that the company requires safety and security practices. In this article, AMODEI also said that the United States should be exported to chips in China to limit the possibility of an undecided global AI race.
Anthropic has always been noticeable in Openai and Google to focus on safety. Other technology companies withdrew the SB 1047, a controversial AI safety bill in California, but issued humble support and recommendations for the bill, which set up a safety report standard for the Frontier AI model developer.
In this case, Anthropic does not seem to increase its function, but seems to be pursuing the industry’s efforts to better understand the AI model.
