AntGroup Open Sources SingProbe: Real-Time Risk Blocking for Large Model Generation with Lightweight Probe, Computing Overhead Below 0.5%

The AI Safety Lab of Ant Group has recently officially open-sourced the intrinsic safety barrier technology for large models, SingProbe. This technology differs from the mainstream external safety review model by reusing the hidden states during the reasoning process of the base model, achieving real-time risk assessment at the token level with less than 0.5% additional computing overhead. It can simultaneously perform intent classification, safety detection, and hallucination identification tasks. It shows significant advantages in high-risk scenarios such as healthcare, enabling millisecond-level blocking before content output.

Currently, SingProbe is compatible with 29 mainstream large models. It is the first complete intrinsic safety technology solution open-sourced by a major tech company. Its open-source code and pre-trained models have been released simultaneously.

GPT-6 Astra Simulates Operating a Vending Machine for a Year and Earns $15,000, Outperforming Claude by Nearly Three Times

In the Vending-Bench2 test by Andon Labs, AI simulated operating a vending machine for a year with a $500 budget. GPT-6 Astra achieved an average account balance of $15,515, ranking first for the first time. Even its worst round exceeded Claude Fable5.1's best performance. The gap is not only in profitability but also in strategy: Astra continuously reduced procurement costs, lowering them to about 48% of the original price at one point.

Huang Renxun: Opposes Slowing Down AI Development, the US Will Continue to Expand AI Infrastructure

NVIDIA CEO Huang Renxun had a live call with U.S. President Trump at the All-In Summit in Los Angeles, where they discussed the pace of AI development and the controversy surrounding data center construction. The conversation was broadcast to the audience via speakers. Previously, the CEO of Anthropic, Amode, had called for slowing down artificial intelligence development, highlighting the industry's division on the speed of AI advancement.

WorkBuddy Upgrades PPT Capabilities: Generate a Presentable Version with One Sentence, Continue Editing the Original File

Tencent's WorkBuddy upgraded its PPT generation on September 14: supports generating a first draft with one sentence, continuing to edit based on the original file, and no need to recreate the entire document. Users can input instructions in the chat window or click on the slide to invoke the "/tencent-pptx" skill. If the materials are incomplete, only the topic and number of pages will be provided, and the system will ask questions about the audience, scenario, and goal, then generate content following the structure of problem - argument - conclusion - implementation. Users can also upload materials such as research reports.

Anthropic CEO Calls for the Entire Industry to Slow Down AI Development, Musk and Altman Both Agree

Dario Amodei, CEO of Anthropic, published an article calling on the front-line AI industry to slow down the pace of technological iteration. He warned that if the development rhythm is not actively constrained, technological breakthroughs will quickly exceed the understanding and control of ordinary humans. After the article was published, Musk also responded publicly.

Anthropic Discloses: Multiple Attempts to Use AI for Developing Biological Weapons Have Been Prevented This Year

According to a report by the Financial Times, Anthropic stated that it has prevented multiple attempts this year where scientists tried to use its AI to potentially develop biological weapons. It listed five cases in which the individuals bypassed controls and misrepresented their intentions to circumvent safety measures. The incidents highlight the potential threats AI poses to public safety.