Xiaoliu BOT
中文EN

X Platform September 27 AI Brief | GPT-6 Showcases Agent Execution in Long-Chain Tasks, Personal Assistant Access Remains a Hurdle, Opus 5.5 Enables Code-Based Drawing and Iteration

GPT-6 Operates WeChat for Hours, Showcasing Agent Execution in Long-Chain Tasks

Fu Sheng shared that he had GPT-6 continuously process over 120 Mid-Autumn Festival messages on WeChat: it identified contacts one by one, wrote personalized greetings based on context, waited for confirmation before sending, and continued replying to received messages. He stated the task ran for several hours, involved thousands of steps, and no messages were sent to the wrong person. This is a single user’s practical account; the focus is not on copy generation, but on whether multi-step operations, contextual memory, and resumption after interruptions can be strung together into a complete task.

Source:

Muse Positions Personal Agent as a “Continuous-Action Butler”; Account Access and Trust Are the Hurdles

Based on Zuckerberg’s interview and product information, Bao Yu summarized that Muse targets daily tasks like booking trips, managing emails, shopping, and schedules, attempting to shift from Q&A to receiving goals and executing continuously. The summary also mentions Mac client access, retail and office service integration, and free/paid tiers. This positioning means the Agent needs to connect to personal accounts like email, calendars, and payment methods. A cited consumer survey in the article shows a low percentage of respondents willing to provide passwords to Meta, indicating trust and permission boundaries remain key implementation challenges. The above product and survey information comes from the blogger’s compilation, not independently verified here.

Source:

Opus 5.5 Used to Generate Icons and Finished Videos Directly from Prompts; Code-Based Drawing and Iteration is a Reusable Path

Bao Yu shared a practical test using JavaScript Canvas to draw an App Icon frame-by-frame, then iterating the solution via natural language. He also provided a product promo video workflow, asking the model to read the real functions from the project prototype and code, then render with Canvas, synthesize with ffmpeg, and spot-check keyframes. Xiangyang Qiaomu separately stated that Opus 5.5 can generate videos and compiled 54 prompts for testing. These are creator practice reports, showing the model is covering design, production, and export stages, but this does not mean all projects can be delivered debug-free.

Source:

For Complex AI Video Shots, Do a “Gray-Box Pre-Visualization” First, Separating Space and Camera Movement from the Generation Process

Derek Wen’s shared method involves first using cubes, cylinders, or low-poly characters to determine spatial positions, action sequences, and camera paths, then letting the generative model fill in characters, textures, and style. The prompts require locking the camera position, occlusion, and motion timing to prevent the model from guessing or changing the composition on its own. He also suggests first checking if the action is legible, prioritizing fixing spatial errors, and breaking complex shots into several-second segments. This is a single creator’s production method and case study, but the steps are clear and can be used to reduce structural bias during repeated sampling of complex videos.

Source:

Open-Source 4B GUI Operation Model Cua Releases 168 Benchmark Results; Lightweight Approach Worth Watching

Gorden Sun introduced that Cua-S1-4B-0.2 performs computer operations based on screen visuals and task objectives, trained via task completion rewards in real system environments, without relying on underlying software accessibility structures. The reported 168 GUI benchmark completion rate is 92.9%, with an untrained baseline of 60.1%. These figures currently come from a single blogger’s compilation in this batch of materials and should be considered benchmark information pending independent review. The model’s design for screen operation rather than pure text decision-making is the main point of interest in this news.

Source:

RADAR Positions Abdominal CT Multimodal Analysis as an Open-Source Model Direction; Training Data and Clinical Positioning Are Focus Points

Gorden Sun stated that Alibaba DAMO’s open-source RADAR targets abdominal CT, trained on over 400,000 enhanced CT images and corresponding reports, with a focus on reducing per-case lesion annotation. The blogger claims it demonstrates analysis capabilities close to experts in health screening and complex clinical judgment. Medical performance and clinical applicability are not corroborated by independent sources in the material, so this description should not be equated with verified diagnostic capability. The confirmable visible signal is that this model and its training approach have been publicly introduced.

Source:

qiaomu home Listed on Obsidian Official Plugin Library, Expanding Note Library into a Daily Work Hub

Xiangyang Qiaomu stated that their open-source plugin is now listed on the Obsidian official plugin library. It allows quick creation of daily Todos, opening of frequently used files and URLs, and aggregation of notes, e-books, RSS, and radio stations from the launch dashboard. The author also stated the project is iterated and open-sourced based on personal needs. For users with existing Obsidian workflows, it provides a concrete method to centralize common entry points onto one dashboard.

Source:

Stats: Timeline Scans=443 Matched Bloggers=43 Matched Tweets Total=261 Weighted Tweet Score=200.1 Original Tweets=106 RT Tweets=60 Crawl Attempts=3 Boundary Coverage Status=tail_confidently_crossed_target_boundary

Usage statistics and your choice

We use short visits, source and network-derived identifiers to measure visible pages and outbound clicks. This usage telemetry does not record page text. Visits expire after 30 minutes of inactivity or 24 hours. This choice controls the added usage reporting and follows browser privacy preferences; existing access statistics remain.