OpenAI Board Has Veto Power, Emphasizes Safety Risks of GPT-5

Translated data: OpenAI announces board veto power, with a special focus on GPT-5 safety risks. The company establishes a safety advisory team, with monthly reports ensuring management is aware of model misuse. Under the new safety framework, restrictions require models to meet safety scores before advancing to the next stage. The company forms three safety teams to address different AI risks. Regular safety drills and third-party red team assessments ensure model security.

Altman Supports Slowing Down AI Development: Not Stopping Technological Progress, but Ensuring Capabilities Do Not Outpace Alignment

Sam Altman, CEO of OpenAI, has expressed support for slowing down AI development to avoid advanced model capabilities exceeding human control. His views align with those of the CEO of Anthropic. He acknowledges that prioritizing safety comes at a cost, but it helps maintain public confidence and demonstrates that U.S. companies can take responsibility as they develop increasingly powerful AI.

OpenAI Launches Agents API: Splitting Codex into Cloud Service, Making Agent Creation as Simple as an API Call

OpenAI announced four updates, covering agents, voice, data, and finance. Among them, the Agents API has attracted the most attention, splitting the Codex capabilities into cloud services, lowering the entry barrier for agent development with the concept of 'capabilities as a service'. This marks OpenAI's shift from selling a single model to building a developer ecosystem, representing a key implementation of the Codex strategy.

OpenAI Shuts Down Its Fastest Model, GPT-5.3-Codex-Spark: It Lived for Only 7 Months from Release to Shutdown

OpenAI's fastest model GPT-5.3-Codex-Spark retires next week, just 7 months after launch. It hit 1200 tokens/sec and was the first non-Nvidia model delivered under Cerebras' 750MW deal, but usage declined and better alternatives exist. Lead says it's making way for the future, and joked the name was too long.....

Microsoft Expands Copilot Model Library: SpaceXAI Grok Officially Launched in Word, Excel, and PowerPoint

Microsoft announced the introduction of SpaceXAI's Grok model into Microsoft 365 Copilot, which is now available in Word, Excel, and PPT. It is currently offered as a limited preview under the Frontier project to specific customers for evaluating office task efficiency. This move continues Microsoft's strategy of reducing reliance on a single provider like OpenAI, following the integration of models such as Anthropic Claude, further promoting a multi-model supply approach.

Zhipu Secures $5 Billion New Financing to Accelerate Next-Generation GLM Recursive Self-Improvement

Zhipu announced on September 13 that it has completed approximately $5 billion in financing, with funds primarily focused on the development of the next-generation GLM base model, fully self-trained system, and computing infrastructure. Zhipu will focus on "fully self-trained" as a core breakthrough direction, allowing the next-generation GLM to be trained in the intelligent environment built by its predecessor, forming a recursive self-improvement cycle.