AGI Lab on Amazon reveals its first job: Advanced AI agents

Amazon is still regarded as a slight lag in the race for the development of advanced artificial intelligenceBut she quietly created a lab that now sets records when it comes to AI. Amazon’s AGI SF LABwhich is located in San Francisco and devoted to construction general intelligenceOr AI, which surpasses people’s opportunities, revealed the first fruits of their work today: a new AI model capable of feeding some of the most modern AI agents available everywhere.

The new model, called Amazon Nova Act, is superior to those from Openai and Anthropic several indicators designed to evaluate the intelligence and ability of AI agents, Amazon says. In Benchmarks Proomui Web and Screenspot, Amazon Nova Act, he performs better than Claude 3.7 Sonnet and Openai Computer Use Agent. A major part of Amazon’s plan to compete in the AI ​​market is to focus on construction agents, and the capabilities of the new model reflect its efforts to build a generation of tools that can measure to the best available.

“I believe that the main nuclear unit of calculations in the future will be a call to a giant agent (AI),” says David Luan, who runs AGI SF Laboratory on Amazon. Previously, he was Vice President of Engineering at Openai, and later complicit ADEPT, a launch, which introduced work on AI agents before joining Amazon in 2024, when the e -commerce giant took a stake in the company.

Most of the leading AI laboratories are already focused on Building increasingly capable AI agentsS Getting AI to master independent actions, as well as conversation, promises to make the technology more useful and valuable. The transition from chat to action is still very underway.

Over the past six months Openai, Anthropic, Google and others demonstrated Web browsing agents who take action in response to a prompt. But for the most part, these agents are still unreliable and can easily be stabbed by open requests.

Luan says Amazon’s goal is to build AI agents that are reliable rather than intrusive. The thing retention of agents is not the need for “cooler demonstrations of interesting capabilities that work 60 percent of the time, but the problem of Waymo,” he says, citing how self-driving cars must be trained to deal with unusual cases before they can go out on the street.

Many so -called agents are built by combining large language models with multiple human -written rules that are intended to prevent them from deviating from the course, but also make their behavior fragile. Amazon Nova Act is a version of the company’s most powerful home model Amazon Nova This has received additional training to help him make decisions on what action to take and at what time. In general, says Luan, AI models are struggling to decide when they should intervene in a task.

To enhance Nova’s agency capabilities, Amazon uses Reinforcementa method that has helped others AI models better simulate reflectionsS