Amazon is still perceived as a bit delayed in the race to develop advanced artificial intelligence, but quietly created a laboratory that now sets records when it comes to AI performance. Amazon Agi sf labwhich is located in San Francisco and devoted to building artificial general intelligence or AI, which exceeds the capabilities of people, revealed the first fruits of their work today: a novel AI model capable of supplying some of the most advanced AI agents available anywhere.
Amazon says that the novel model, entitled Amazon Nova Act, exceeds those from OpenAI and Antropic on several comparative tests designed to assess the intelligence and skills of AI agents. At Benchmarks Granundi Web and Screenspot Amazon Nova Act works better than Claude 3.7 Sonnet and OpenAi Computer Agent. The main part of the Amazon plan for competition on the AI market is to focus on construction agents, and the skills of the novel model reflect his efforts to build the generation of tools that can measure the best available.
“I think that the basic atomic unit of calculations in the future will be a call to the giant [AI] Agent, ”says David Luan, who runs the Aga SF Amazon laboratory. Earlier he was the Vice President of Engineering at Openai, and later a commercially co -founded the adept, a startup who would be a pioneer in AI agents, before he joined Amazon in 2024, when the E -Commerce giant took part in the company.
Most of the leading AI laboratories are currently focusing on building more and more talented AI agents. The impediment to artificial intelligence to master independent actions, as well as conversations, promises that technology will be more useful and valuable. However, the transition from chat to action is still very undergoing.
Over the past six months, Opeli, Anthropic, Google and others have demonstrated web browsing agents who take action in response to prompting. But for the most part, these agents are still unreliable and can be easily stumbled with the help of open conclusions.
Luan says that Amazon’s goal is to build AI agents who are reliable than flashy. The thing stopped by agents is not the need for “more frigid demonstrations of captivating possibilities that work in 60 percent of cases, this is a problem with Waymo,” he says, referring to how self -propelled cars should be trained to deal with unusual matters before they could choose streets without supervision.
Many so -called agents are built by combining vast language models with many principles written by people designed to prevent them away, but also makes their behavior feeble. Amazon Nova Act is a version of the company’s most powerful home Amazon Nova This received additional training that will assist you decide what actions should be taken and at what time. In general, Luan says that AI models are trying to decide when they should intervene in the task.
To improve Nova’s agency skills, Amazon uses reinforcement learning, a method that has helped other AI models better simulate reasoning.