Amazon is still seen as a bit behind in developing advanced games AIbut it quietly created a lab that is now setting records when setting up AI performance. Amazon’s AGI SF Laboratorylocated in San Francisco, dedicated to construction Artificial General Intelligenceor AI that exceeds human capabilities, reveals the first fruits of today’s work: a new AI model that powers some of the most advanced AI agents available anywhere.
Amazon says the new model, called Amazon Nova Act, outperforms several benchmarks for Openai and anthropomorphism, aims to evaluate the intelligence and capabilities of AI agents. Amazon Nova Act performs better on benchmark Groundui networks and ScreenPot than using proxy for Claude 3.7 sonnets and OpenAI computers. An important part of Amazon’s plan to compete in the AI market is focusing on construction agents, and the capabilities of the new model reflect its efforts to build a generation of tools that can achieve optimal availability.
“I believe that the basic atomic units of the computer in the future will be a call to the giants [AI] Agent.
Now, most of the leading AI labs are focused on Build increasingly capable AI agents. Letting AI master independent actions and dialogue will hopefully make technology more useful and valuable. However, the transition from chat to action is still underway.
Over the past six months, Openai, Anthropic, Google and others have proved Web browsing agent This takes action to respond to the prompt. But in most cases, these agents are still unreliable and can easily trip over them with open requests.
Luan said Amazon’s goal is to build reliable rather than flashy AI agents. He said that the thing that blocks agents doesn’t require “a cooler demonstration of fun features working 60% of the time, which is a Waymo issue”, referring to how training with impatient conditions requires training for autonomous cars to handle abnormal edge cases.
Many so-called agents are constructed by combining large language models with multiple human-written rules designed to prevent them from deviating from the route, but also make their behavior brittle. Amazon Nova Act is the company’s most powerful version of native models Amazon Nova This has received additional training to help it decide what to take and at what to take. Often, AI models work hard to decide when tasks should be intervened, Luan said.
To improve NOVA’s applicability, Amazon is using Reinforcement learninga way to help others AI models better simulate reasoning.
