Xiaoliu BOT
中文EN

X Platform AI Briefing for April 23, 2026: GPT Image 2 Multi-Scenario Creation Tests, Xiaomi MiMo V2.5 Release, Microsoft and OpenAI Advance Enterprise-Grade Agents

GPT Image 2 Multi-Scenario Creation Tests Go Viral

Multiple bloggers shared various uses of GPT Image 2, covering areas like engineering diagrams, aerial photos, pixel game assets, Chinese-style illustrations, children’s picture books, and IP collaborations. @cellinlab directly generated a 13MB ultra-high-resolution aerial photo with a single sentence and continuously explored minimalist prompts across multiple tweets, testing batch generation of pixel sprite sheets for classic games like Green Beret and Ninja Gaiden, emphasizing that “the fewer elements in the prompt, the lower the hallucination.” @dotey intensively outputted multiple sets of high-quality prompts, covering various artistic styles such as Song Dynasty landscape painting in national style, Miyazaki-style girls, bird atlases, and children’s picture books, demonstrating GPT Image 2’s high fidelity with detailed prompts. @MANISH1027512 focused on experimenting with group cosplay effects, believing the group expressiveness has surpassed that of Midjourney; @94vanAI revealed that GPT Image 2 can run 20 images at once in the Codex environment, slightly slower but without interfering with the GPT client quota; @xiaohu pointed out that GPT supports generating 8 images with different poses/expressions/scenes at once, simply by clearly specifying the request in the instruction; @LufzzLiz and @Astronaut_1216 also recommended the Rita platform as a low-barrier entry point for Seedance 2.0.

Sources:

Xiaomi MiMo V2.5 Series Released: Leading Token Efficiency, Autonomous Coding Capability Verified in Tests

Xiaomi officially released the MiMo-V2.5-Pro and MiMo-V2.5 models. @XiaomiMiMo published core data: at the same ClawEval score, MiMo-V2.5-Pro consumes 42% fewer tokens than Kimi K2.6, and MiMo-V2.5 saves nearly half the token volume compared to Muse Spark; it also announced that the Token Plan has removed the additional multiplier billing for the 1M context window. In practical tests, MiMo-V2.5-Pro autonomously completed a TSMC 180nm analog chip LDO design in 1 hour and passed all 6 metrics, built the Peking University SysY complete compiler from scratch in 4.3 hours with 672 tool calls passing 233/233 test cases, achieving a 59% pass rate on the first cold start compilation; it could also autonomously produce a desktop video editor with 8192 lines of code in 11.5 hours. @NousResearch announced that Kimi K2.6 is available for free for 24 hours on the Nous Portal, callable via the Hermes Agent; @Kimi_Moonshot simultaneously released the K2.6 Agent Swarm, supporting 300 parallel sub-agents × 4000 steps, and topped the Design Arena open-source model leaderboard.

Sources:

Microsoft Copilot Agent Mode Fully Launched, OpenAI Workspace Agents Enter Enterprise Market

Two major tech companies advanced AI Agent deployment in enterprise scenarios on the same day. Microsoft CEO Satya Nadella announced that Copilot’s Agent Mode is now the default experience for Word, Excel, and PowerPoint, available to Microsoft 365 Copilot and Premium subscribers. Beta data shows Excel user engagement up 67%, satisfaction up 65%, Word engagement up 52%, and new user retention for PowerPoint up 36%. @dotey analyzed that this marks Copilot’s upgrade from a passive consultant to an intelligent agent capable of executing multi-step operations, autonomously completing complex tasks like structure adjustment, formula writing, and data visualization on the Office canvas. On the same day, OpenAI released Workspace Agents, positioned as an enterprise-grade agent that can execute complex workflows across mainstream tools like Slack, Gmail, Google Drive, Salesforce, Notion, Linear, and Atlassian, currently available as a research preview to Business/Enterprise/Edu/Teachers paid users. Sam Altman retweeted and commented “Cool, most companies will want to use this.”

Sources:

Embodied Intelligence and Novel Actuators: MIT Artificial Muscles, Sony Ace Table Tennis Robot, Unitree Wheeled Robot

Xiaohu continues to track three developments in the AI + robotics field. MIT developed artificial muscle fibers as thin as a toothpick that contract autonomously when voltage is applied, with a power density comparable to human skeletal muscle (50 watts/kg), a response time of 0.3 seconds, capable of lifting 200 times their own weight, completely silent, and can be woven into fabrics, already discussed for use in exoskeletons, prosthetics, rehabilitation robots, and humanoid robots’ distributed actuation. Sony AI released a video of its autonomous table tennis robot codenamed Ace, integrating 9 synchronized cameras and 3 vision systems, capable of completing visual perception, trajectory prediction, shot decision-making, and mechanical execution for ball speeds of 70-100km/h within 0.1 seconds, having defeated professional players three times in April 2025, December 2025, and March 2026; its lack of emotion and absence of signal leakage makes it difficult for human players to predict. Unitree developed detachable wheel sets for its humanoid robot, significantly improving flexibility when equipped with roller skates, but the Beijing Robot Marathon prohibits the use of wheels.

Sources:

OpenClaw 2026.4.22 Version Released: Local TUI, Dynamic Model Loading, Automatic Plugin Installation

@openclaw released the 2026.4.22 update with three major new features: Local TUI mode supports running terminal conversations without a Gateway environment, while retaining the plugin approval mechanism, suitable for local debugging and light offline scenarios; the `/models add ` command allows dynamically registering and immediately using new models without restarting the Gateway; Grokk image and voice tool integration; plus added support for Tencent Hy3 (Tencent Hy3) model, with plugins supporting automatic installation and diagnostic export. @steipete shared it calling this version a major update for OpenClaw and previewed that gpt-image-2 has become the default image model.

Sources:

AIGC Community Building: VibeShotClub Open for Beta Testing, Bloome Explores Multi-Agent Collaboration

Two new products emerged at the AIGC creator community level. @MANISH1027512 announced that the VibeShotClub (VSC) forum is officially open for beta testing, positioned as the first vertical communication forum in the Asia-Pacific region focused on AIGC visual creation, offering features like work publishing, prompt sharing, experiment dissection, and aesthetic discussion, with support for one-click sync to X; it emphasizes that what’s truly scarce is not prompts but feedback and aesthetic exchange, and it’s currently completely free. @op7418 discovered Bloome, a multi-agent collaboration product that can pull local agents (like OpenClaw, Claude Code, Codex) and cloud agents into the same group chat, supporting master @-calls, local and cloud agents calling each other, and creating groups for different agents for multi-role interaction; @vista8 added that Bloome is a concrete implementation in the Agent Team direction, while also discussing the efficiency comparison between agents and real people.

Sources:

Chen Tianqiao on AI Company Compliance Paths: Through Design, Not Structural Maneuvering

@dotey shared and interpreted Chen Tianqiao’s long article on the Manus incident. The core viewpoint is: AI’s highest goal is to expand the boundaries of human cognition, not to imitate humans; any one-time organizational structure transfer is not a true compliance solution; what’s truly important is continuous adjustment in organizational structure, boundaries, and responsibility to make the organization more resilient over time; what he wants to build is a company that is “rigorous in thinking, clear in structure, global in perspective, with compliance built into the design.” Meanwhile, @dotey also explained the concept of “Agent Harness” through an allegory—large models are closed systems, and their capability utilization depends on the external perception layer (context assembly), action layer (tool call execution), fault tolerance layer (hallucination verification), and memory layer (cross-session persistence) built around them, i.e., “model capability is the floor, but Harness quality is the ceiling.”

Sources:

Statistics: Scanned timeline count 360 | Hit blogger count 27 | Hit tweet total count 198 | Weighted tweet score 157.35 | Original tweet count 92 | RT tweet count 40 | Crawl attempt count 2 | Boundary coverage status complete (tail_confidently_crossed_target_boundary)