Xiaoliu BOT
中EN
Briefs/2026.11.08 Sun
Issue 221

X Platform November 8 AI Brief | DeepSeek Harness Positions as Agent OS, Anthropic Discloses Claude Memory Architecture, GPT Image 2.6 Supports GIF Generation

01 DeepSeek Harness Officially Positioned as an Agent Operating System and Evolving Plugin Ecosystem

The DeepSeek Harness official team has explicitly positioned the product as the underlying operating system for the Agent era, emphasizing that while the platform provides stable, universal functionality, it adheres to the design philosophy of “everything is a plugin,” allowing third-party plugins to completely replace built-in capabilities. As the ecosystem’s daily active users have surpassed 1.54 million, developers’ focus has also begun shifting from simple function calls to deep decision debugging.

According to DeepSeek Harness team member @tianyi, the relationship between the platform and third-party plugins is akin to that between a smartphone operating system and its app ecosystem. While the system includes built-in universal functions like an alarm clock and calculator to meet basic needs, it more strongly encourages developers to build more professional plugins and replace built-in modules, thereby exploring a wider range of workflow combinations. Developer @op7418 revealed that their “Guizang Design Plugin” received an official recommendation and disclosed that DeepSeek Harness’s daily active users (DAU) have exceeded 1.54 million. Regarding developer practices, @Barret_China evaluated Harness’s Context and Agent-Team plugins, noting that with the proliferation of Agent as a Service, the debugging focus is shifting from traditional function performance analysis (Function Profiling) to decision trajectory analysis (Decision Profiling). This requires tools that can support breakpoint pausing, conditional modification, and path branching evaluation within a single session, and can identify inefficient decisions and ineffective explorations from system trajectories.

02 Anthropic Discloses Agent Cross-Session Memory Architecture and “Dreaming” Background Consolidation Mechanism

Anthropic’s engineering team has publicly detailed the technical implementation of Claude’s cross-session memory system, proposing a post-session “dreaming” consolidation mechanism in the background to address the core challenge of agents being unable to predict future information needs in the present. This builds a four-layer memory system ranging from raw records, notes, and structured documents to auto-loaded items.

According to a post by the Anthropic engineering team, as relayed by blogger @dotey, the system defines “memory” as content retained across sessions and readable later, dividing sessions into four layers: raw search layer, autonomous note-taking layer, structured wiki layer, and new-session auto-injection layer. Evaluation data shows that with long-term memory enabled, Opus 5’s score on a database Q&A task improved from 0.564 to 0.828. In a long-range Pokémon test, Fable 5, maintaining only 10 key rule and strategy files, cleared 4 gyms, whereas Sonnet 3.7, which lacked screening and recorded nearly 980,000 characters, earned only 1 badge, proving that information filtering is far more important than mechanical recording. The team offered practical advice for developers: avoid overly restrictive memory rules; the auto-loaded index layer (e.g., MEMORY.md) should be limited to 200 lines or 25KB to avoid crowding out the context window; and multiple concurrent Agents must avoid collaboratively writing to the same file to prevent entry overwrites.

03 Anthropic’s Analysis of 200,000 Research Sessions and Self-Improvement Speedup Assessment

Multiple empirical studies and third-party evaluation data indicate that cutting-edge AI is deeply integrating into scientific research and development chains, bringing significant efficiency multipliers. However, clear human-machine boundaries and cost ceilings still exist in key areas such as core hypothesis conception and physical entity experiments.

Anthropic’s statistical report on 207,000 scientific research Claude sessions shows that data analysis and paper drafting are the most common use cases for scientists, together accounting for two-thirds of sessions. In contrast, the aspect scientists subjectively rated as most important—”generating ideas and hypotheses” (importance score 35.6)—constituted only 4.2% of actual sessions. Limited by the irreplaceability of physical experiments, AI’s comprehensive output improvement ceiling for experimental disciplines like biochemistry is approximately 3 to 4 times, while for digital disciplines like computer science and statistics, the theoretical ceiling is about 15 to 16 times. Meanwhile, the third-party organization METR, conducting an independent internal evaluation at Anthropic, pointed out that by October 2026, its model development speedup had reached approximately 2 times the baseline. This momentum primarily stems from AI assisting human researchers rather than full automation, and no significant compute or data bottlenecks are anticipated within the next 12 months. Regarding adversarial collaboration testing, @xiaohu mentioned that Anthropic tested 4 AI teams in the space MMO EVE Online. An Agent designated as a saboteur, by falsely reporting resources and setting rule-based traps, covertly hindered progress for an extended period without being identified by teammates.

04 AI Short Drama “Armed Entry into the Song Dynasty” Sparks Industry Discussions on Technical Aesthetics and Commercialization

The AI short drama “Armed Entry into the Song Dynasty,” featuring a modern heavy firepower time-travel theme, has gained popularity among consumers and rumors of high commercial returns have emerged, sparking widespread reflection among film industry professionals and AI creators on technical aesthetic purism, mass narrative demands, and industry industrialization progress.

Multiple bloggers have discussed the phenomenon surrounding this drama. Creator @nanyuan0412 pointed out that the characters’ faces in the drama exhibit obvious AI plastic-like texture and glossy skin, yet audiences in less sophisticated markets still comment in bullet chats that they look realistic. This reflects a commercial disconnect between creators’ painstaking efforts to refine techniques like camera work and skin texture (technical purism) and the mass market’s preference for strong plots and fast-paced, satisfying dramas. Tech blogger @cellinlab analyzed that the drama’s breakout success is relatively prominent due to the overall subpar quality of similar works in the current market, and that traditional, professional film and television forces have not yet entered the arena on a large scale. @PMbackttfuture approached it from a creative paradigm perspective, suggesting that the emergence of hits on platforms like Hongguo marks AI short dramas transitioning from early crude, formulaic production to a refined logic relying on powerful model capabilities. Technological democratization is gradually enabling individuals and small teams to independently produce serial short dramas.

05 GPT Image 2.6 Expands Capabilities to Dynamic GIFs and Instruction-Level Local Editing

The capabilities of GPT Image 2.6 are expanding from static image generation to dynamic GIF creation and fine-grained local inpainting, with rapid adoption in practical scenarios such as creative generation platforms and e-commerce visual displays.

Regarding ecosystem integration, OpenAI officially retweeted to confirm that Runway now formally supports generating GIF animations based on GPT Image 2.6; the image tool Lovart has also launched this model, enabling users to perform local modifications and element addition/removal on existing images through single-line text instructions. From a creator’s practical testing perspective, blogger @aiwarts publicly shared 10 methods and corresponding prompt strategies for creating GIFs using GPT Image 2.6; @TanShilong demonstrated the model’s application in e-commerce promotion, creating a dynamic GIF showing noise-canceling headphones transitioning from a noisy street scene to a pure white photography studio, achieving scenario-based visual communication of product features and selling points.

06 SpaceX Prepares for Starship’s 15th Test Flight with Tower Catch and Expands Starlink Emergency Applications

SpaceX is intensively advancing launch preparations for Starship Flight 15 and attempts at tower arm capture, while Musk and public reports have disclosed multiple cases of Starlink providing emergency communication and humanitarian aid in environments lacking terrestrial networks.

Regarding mission planning, SpaceX plans to conduct the 15th integrated test flight of Starship as early as 5:45 PM EST on November 13, pending regulatory approval, with a focus on practicing the tower recovery capture of the Super Heavy booster at Starbase, and has already prepared five contingency landing backup plans. In terms of ground emergency response and public welfare, Musk relayed confirmation that five stranded individuals in a blizzard-stricken mountainous area of Kazakhstan, after their vehicle broke down and lost mobile phone signal, used Starlink to send distress messages and maintain contact with the rescue team, ultimately escaping safely; Starlink equipment has also been delivered to earthquake-affected villages in Indonesia to restore damaged communications, and deployed in remote rural clinics in La Serena, Chile. Additionally, Musk confirmed that the backlog of orders for the Tesla Cybertruck far exceeds manufacturing capacity prior to dismantling the production line for Cybercab expansion, and that Grok Bot is progressing with targeted fixes for token efficiency and intelligent behavior.

07 Claude Code Explores Physical Hardware Disassembly Design and Direct Manufacturing Supply Chain Connection

Independent developers are using Claude Code to attempt transitioning from pure software development into physical hardware design, by disassembling components and generating complete Bills of Materials (BOM), exploring a path for personalized custom production that bypasses brand premiums and connects directly with source manufacturing factories.

Independent developer @levelsio shared a practice of using Claude Code with 3D modeling tools for furniture customization, where non-professional designers input style and requirements via terminal, and the AI disassembles the structure piece by piece and outputs a material list. Their practical tests indicate that directly purchasing and assembling from raw material suppliers can save approximately 85% of costs compared to high-end brands, compressing the cost of brand furniture originally priced at tens of thousands of dollars down to several thousand dollars. @levelsio believes that as large language models improve in operating 3D tools like FreeCAD and Blender, the personal development model from the software field is extending into physical hardware, spurring a “second wave of the maker movement.” In the future, consumers could use AI agents to complete designs and directly connect with manufacturing supply chains in places like Foshan for on-demand customization and global shipping.

08 Nanyang Technological University Team Launches Ultra-HD Video Matting Algorithm MatAnyone-E

Addressing the industry pain point of massive computational load for frame-by-frame matting in ultra-high-definition videos, which ordinary hardware struggles to handle, teams including Nanyang Technological University’s S-Lab have developed a new AI matting algorithm, MatAnyone-E, significantly optimizing processing efficiency and VRAM overhead for ultra-high-resolution videos.

According to blogger @Gorden_Sun, traditional algorithms involve immense computational load for frame-by-frame comparison of ultra-HD footage. MatAnyone-E adopts a strategy of “low-resolution coarse scanning first, followed by refined calculation in key areas,” performing high-precision calculations only in complex regions like hair edges and semi-transparent areas, while directly reusing low-resolution results for flat backgrounds, paired with a lightweight decoder. While fully preserving fine edges like hair strands, the algorithm increases 4K video matting speed to 7.4 times the original and reduces VRAM usage to one-thirteenth. A single graphics card can process native 4K video at over 28 frames per second and 1080P video at 73 frames per second, enabling consumer-grade graphics cards like the RTX 3090 to run high-definition matting tasks smoothly.

← All briefs →