Amazon’s Agi Laboratory reveals his first work: advanced artificial intelligence agents

Amazon is still seen as a backward in the race to develop the applicant artificial intelligenceBut it was created quietly a laboratory that now prepares records when it comes to performing artificial intelligence. Amazon AGI SF LABWhich is located in San Francisco and devoted to construction Artificial general intelligenceOr AI, which exceeds the capabilities of humans, revealed the first fruits of its work today: a new model for Amnesty International capable of operating some of the most advanced artificial intelligence agents anywhere.

Amazon says the new model, called Amazon Nova Act, is outperforming the performance of Openai and Anthropor in many criteria designed to measure the intelligence of artificial intelligence agents. On Groundui Web and Screenspot standards, Amazon Nova Act performs better than Claude 3.7 Sonnet and Openai Computer Use Agent. A large part of the Amazon plan is to compete in the artificial intelligence market to focus on construction agents, and the capabilities of the new model reflect its efforts to build a generation of tools that can reach the best available.

“I think the basic atomic unit of computing in the future will be an invitation to a giant agent (AI),” says David Luan, who leads the AGI SF laboratory at AGI SF. He was previously vice president of engineering at Openai, and then, he was postponed Adept, a leading startup company in AI’s customers, before joining Amazon in 2024 when the e -commerce giant took a stake in the company.

Most of the leading AI laboratories are now focusing on Building artificial intelligence agents increasingly. Obtaining Amnesty International to master independent procedures, as well as conversation, is to make technology more beneficial and valuable. However, the shift from chatting to work is still ongoing.

In the past six months, Openai, Anthropic, Google, and others have shown Web browsing agents That takes action in response to a claim. But for the largest part, these agents are still unreliable, and they can be easily stumbled through open requests.

Luan says that Amazon’s goal is to build a reliable Amnesty International agents instead of brightness. The thing that hinders agents is not the need for “more illustrative offers with interesting capabilities that operate by 60 percent of the time, it’s a Waymo problem”, referring to how self -driving cars are training in dealing with unusual edge cases before you can go to the streets not subject to supervision.

Many alleged factors are built by combining large language models with multiple human written rules that are designed to prevent them from deviating from the path, but also makes their behavior fragile. Amazon Nova Act is a copy of the most powerful local model for the company Amazon Nova This has received additional training to help him make decisions on the measures to be taken at any time. In general, Luan says, Amnesty International models are struggling to determine when to interfere with a mission.

To improve Nova Agential capabilities, Amazon is used Learning reinforcementHow to help others Artificial intelligence models better thinking.