Video Conferencing Giant Zoom Surpasses Records with Federated AI in the World's Toughest AI Exam

For a long time, the "Wudang Summit" in the field of large models has been dominated by laboratories such as Google, OpenAI, and Anthropic. However, last week, video conferencing giant Zoom dropped a shocker: its artificial intelligence system achieved a score of 48.1% in the top benchmark test called "Humanity's Last Exam (HLE)", breaking the world record and surpassing the previous record held by Google Gemini3Pro at 45.8%.

Meeting, Meeting

Not Training Models, but "Tactics": The Federated AI Approach

Facing the question of "how a non-core model vendor can overtake model giants," Zoom's Chief Technology Officer and former Microsoft AI leader Xuedong Huang provided a completely different answer. Instead of spending money to train its own trillion-parameter large model, Zoom developed a precise "Federated AI approach":

Z-Scorer: The core brain of the system, responsible for real-time evaluation of responses from multiple models such as OpenAI, Google, and Anthropic, selecting the optimal solution for specific tasks.
Explore-Validate-Union Strategy: This is an agent workflow that enables "dialectical collaboration" among multiple AI systems, challenging and refining reasoning results together.
Traffic Controller: Simply put, Zoom built an extremely smart AI "dispatch center" that achieves performance beyond any single model by integrating rather than self-developing.

Controversy: Is It Real Innovation or "Stealing Home"?

This achievement sparked significant debate within the developer community. Critics like engineer Max Rumpf argue that Zoom merely "wrapped" others' achievements through APIs, scoring high in benchmark tests with limited practical significance for actual users, calling this behavior "plagiarizing others' labor" without substantial technical contributions.

However, other observers such as developer Zhu Hongcheng believe that winning in AI evaluation necessarily requires a "model federation." This is similar to how winners on data competition platform Kaggle always rely on model ensembles rather than going it alone. This strategy is extremely clever in business terms: it avoids expensive computing power investments while allowing Zoom to flexibly switch between suppliers, completely avoiding supplier lock-in.

From the Rankings to the Product: The Practical Test of AI Companion 3.0

Xuedong Huang sees this achievement as validation of Zoom's strategy. For Zoom's 300 million users, more meaningful competition will take place in the upcoming AI Companion 3.0

Surge in Demand for Humanoid Robot Joints: Global Actuator Market May Approach $10 Billion by 2031

The global market for humanoid robot actuators is experiencing explosive growth. As the "joints" and "muscles" of robots, actuators are the core hardware that enable physical movement, responsible for converting software instructions into precise actions, directly affecting the robot's flexibility, balance, and strength. According to the latest report, the market size is expected to surge from about $1.5 billion in 2024 to nearly $9.9 billion by 2031, showcasing significant commercial potential.

Order Tea with One Sentence, Tongyi App Completes the AI Payment Loop, Alipay AI Pay Enables Intelligent Agents to Actually Get Things Done

The Tongyi Qianwen App integrates with Taobao Flash Sales and Alipay AI Pay. Users can complete product recommendations, orders, and payments through natural conversation without switching apps, achieving 'say a sentence and it's delivered'. This marks a key advancement in AI intelligent agents moving from 'being able to answer' to 'being able to actually get things done'.

Video compression rate reaches 0.02%: China Telecom launches generative video compression technology GVC

The China Telecom AI Research Institute has launched the GVC generative video compression technology, which can achieve a compression rate of up to 0.02%. A 1GB video only needs to transmit about 200KB of data to restore clear image quality. Its core is to use computing power to save bandwidth, differing from the traditional encoding logic of 'moving pixels'. By using an AI model to learn video content features, it intelligently reconstructs the image on the receiving end, significantly reducing the amount of transmitted data, and has the potential to revolutionize video storage and streaming media transmission.

Deep in the AI Pornography Controversy: 28 Organizations Write a Joint Letter, Demanding Removal of X and Grok

Musk's AI assistant Grok is accused of generating sexualized deepfake content involving real people and minors, causing controversy. Twenty-eight American social organizations wrote a joint letter to Apple and Google CEOs, demanding the removal of the X platform and Grok app, accusing the two companies of not strictly enforcing app store policies.