OpenAI's Astra Demonstrates AGI-Like Capabilities in Private Demonstration

ME AI News, Dongcha Beating AI Brief: OpenAI recently hosted a private preview of Astra, inviting executives from dozens of key clients to its newly opened MB0 office in San Francisco. Altman had just returned from Washington, where he had previously given closed-door briefings on Astra to U.S. officials. Outside, protesters held signs demanding “Stop the AI Race,” and OpenAI even brought in an entire wall of greenery to block the demonstrators from view through the glass doors. The demonstration showcased two primary capabilities. First, a research-level math problem: 16 AI agents autonomously broke the problem into multiple subtasks, solved them individually, coordinated with each other, and ultimately assembled a candidate proof. Second, direct computer operation: Astra rapidly switched between common desktop applications, creating and modifying content on its own. Altman said that watching Astra operate the computer in a “superhuman, extremely fast” manner was one of the most striking experiences for OpenAI employees recently. He added that Astra will enable persistent agents—virtual colleagues capable of working continuously over long periods. Astra has also met OpenAI’s internal standard for an “AI research intern.” Given an experimental idea, it can independently implement it in OpenAI’s codebase, run experiments, and return results; given a paper, it can complete work that previously required a human researcher an entire week. Altman believes Astra’s most significant capability is discovering new knowledge. In front of these clients, he stated: “I expect this will be the first model to genuinely invent something meaningful.” He then added: “This is very much like AGI.” (Source: BlockBeats)