Xiaoliu BOT
中EN
Briefs/2026.11.12 Thu
Issue 225

X Platform November 12 AI Brief | NVIDIA CEO Jensen Huang Steps Down; Google Launches Gemini Decisions API and Gemini Robotics 2.5

01 NVIDIA Officially Announces Jensen Huang’s Departure, Interim CEO and Succession Process Clarified

NVIDIA has officially released a CEO transition announcement. Jensen Huang has resigned from his positions as President, CEO, and member of the Board of Directors, effective immediately. The company has appointed Executive Vice President and CFO Colette Kress as Interim CEO and will initiate a formal search for a permanent successor, considering both internal and external candidates. Subsequently, various succession rumors and parody content have appeared on social platforms, but these should not be confused with the official announcement; currently, the confirmed transition information is based on the company’s official statement.

02 Gemini Decisions API Converts Multimodal Input into Batch Structured Judgments

Google has launched the Gemini Decisions API, powered by Gemini Decision 1, designed for closed-option judgments rather than token-by-token text generation. Google states it can evaluate up to 1,000 questions at once on text, images, audio, video, and PDFs, supporting yes/no questions, multiple-choice with up to 255 options, and rubric scoring. Input pricing is $0.10 per million tokens, with no charge for output tokens. Google gives an example where the model processes 100 delivery photos in 4.5 seconds, answering 16 inspection questions per photo simultaneously, making it suitable for scenarios requiring batch screening and triggering subsequent workflows. These performance figures are as disclosed by the publisher.

03 Gemini Robotics 2.5 Begins Rollout to Partners, Official Simulation Training and Test Results Disclosed

Google has released the Gemini Robotics 2.5 vision-language-action model and states it has begun rolling it out to robotics partners. Google introduces Gemini 4 Argon as the simulation engine for teaching new skills that typically rely on large amounts of physical data. In a 200-task evaluation, the model showed a 30% to 100% improvement in success rate compared to previous baselines, depending on task type, and was 20% to 30% faster. These are internal evaluation results published by Google. Task examples include delivering food with a clamp, tying a gift box bow, and pipetting, and should not be considered independent third-party verification.

04 AlphaGenome 2 Opens Non-Commercial Research API, Focusing on Cell-Type-Level Variant Effects

Google DeepMind lead Sundar Pichai announced that AlphaGenome 2 is now available via API for non-commercial research. It is described as a model for predicting the effects of DNA variants across different cell types, helping researchers link genetic discoveries to specific cellular mechanisms of disease. This is an announcement of availability for research purposes; the source did not provide independent evaluation or commercial availability information.

05 Doubao Desktop Version Integrates Third-Party Models: A User’s Practical Test of Video Editing Cost and Time

AI product blogger Guizang stated that the Doubao Work desktop version has added third-party model integration, listing options like DeepSeek, Kimi, and Zhipu GLM, and mentioned that simple tasks in automatic mode are free, with some models offering time-limited discounts. Using DeepSeek V4.1 Flash with his own video editing skill, he processed an 8-minute video, claiming it took about an hour to complete voiceover editing, packaging, and splicing/rendering, with points converted to approximately ¥0.12. This is a personal test and cost estimate and does not represent official pricing or universal performance. His workflow includes clarifying requirements, researching tools, proceeding with production, self-checking, and solidifying the experience into a skill.

06 Grok Bot 1.0 Positions Itself as a Team Collaboration Agent, Integrates with Gopuff for Ordering and Delivery

Grok Bot officially released version 1.0, stating that teams can share Bots, collaborate in Slack and Teams, invite Bots to meetings, and enable continuous learning. The official statement also claims Bots are already handling outcomes in accounting, customer service, sales, and marketing. Another visible integration is with Gopuff: Gopuff states users can purchase its products within Grok Bot for delivery, with the fastest delivery in 15 minutes. The available materials are product-side release information and do not provide independent usage reviews.

07 Fable 5.2 Case Study on Creating a Web 3D Action Game Demonstrates Long-Process Agent Workflow

AI product blogger Guizang demonstrated a case study of creating a 3D action game using Fable 5.2, stating that models, scenes, actions, and optimization were all handled by the tool, with the result integrated into an approximately 30MB HTML file. He later clarified it was not completed with a single prompt but involved multiple rounds of revision, leveraging AI game-making experience and creating a dedicated skill. This skill first inquires about requirements, researches tools, then proceeds with visuals and performs self-checks and verification, also including a checklist for common pitfalls like file size compression. The case is based on the author’s description and demonstration and does not mean similar tasks can be replicated in the same timeframe or cost.

← All briefs →