Xiaoliu BOT
中文EN

X Platform August 31 AI Brief | DeepSeek Multimodal Model Open-Sourced, ChatGPT Ad Revenue Annualizes at $10B, Zhipu API Revenue Surges

DeepSeek Releases First V4 Multimodal Experimental Model with Open Weights and Inference Components

Multiple bloggers have documented that DeepSeek has launched DeepSeek-V4-Flash-Vision-Exp on Hugging Face under an MIT License. Building upon V4-Flash, it adds a vision encoder and Aligner, enabling the model to process images, webpage screenshots, software interfaces, and charts while retaining reasoning, tool calling, and Agent capabilities. The post also lists several benchmark improvements: Terminal Bench 2.1 increased from 82.7 to 83.9, DeepSWE from 54.4 to 59.3, and Toolathlon-Verified from 70.3 to 75.9; specific scores for vision-related tests were also provided. The release includes not only the model weights but also the Tokenizer, Prompt Encoding, a minimal PyTorch inference implementation, and components like the Vision Encoder and MoE. The key focus is on directly integrating multimodal capabilities into Agent workflows.

Sources:

DeepSeek Harness Shifts Towards a Unified Agent Runtime Architecture

According to @MaxForAI’s summary of version changes, DeepSeek Harness released v0.1.2-alpha.1 on August 27th and alpha.2 on August 30th, with core changes involving underlying protocols rather than ordinary feature additions. The old APIProxy was removed, plugin communication migrated to Remote Gateway, the web client was restructured, and Sessions are now handled as replayable event streams. Field changes in alpha.1 introduced compatibility and session recovery risks for some older plugins, which alpha.2 addressed by restoring relevant fields and adding RemoteError, reconnection logic, and plugin scopes. Another direction is that Subagents can now be independently configured with model providers, models, reasoning intensity, and output length. This indicates Harness is integrating models, Providers, sub-Agents, Sessions, plugins, permissions, and communication into a unified runtime, though the version remains in a phase of rapid iteration.

Sources:

ChatGPT’s Ad Monetization and Codex Usage Scale Rise Simultaneously

According to @LufzzLiz citing OpenAI disclosures, ChatGPT Ads, launched less than 200 days ago, has reached an annualized revenue run rate of $1 billion at its current pace. It already has tens of thousands of advertisers, covers over 40 countries, and supports CPC, performance-optimized bidding, Pixel, Conversions API, product feeds, and audience targeting. Its key asset is users’ actively expressed goals and constraints within conversations. However, whether ads influence responses and whether advertisers can access private conversations remain trust boundary issues. Concurrently, @thsottiaux reported that ChatGPT Work and Codex have reached 25 million active users, with paid subscriptions resetting usage limits. The former shows conversation interfaces beginning to carry performance ads, while the latter indicates that the usage scale of Agent/work products continues to expand.

Sources:

Zhipu Open Platform Revenue Surges, But Losses Not Yet Reversed

According to @LufzzLiz’s analysis based on Zhipu’s Hong Kong Stock Exchange interim results announcement, Zhipu’s open platform and API revenue for the first half of the year increased from 29.1 million yuan to 825 million yuan, accounting for 86.5% of total revenue. The gross margin for this business also rose from -0.4% in the same period last year to 24.6%, while the proportion of revenue from on-premises deployments decreased to 13.5%. This indicates a shift in commercialization focus from project-based or on-premises deployments to ongoing API calls, with API gross margins turning positive. However, the company’s total revenue for the first half was 954 million yuan, R&D expenses were 2.131 billion yuan, and operating losses were 2.147 billion yuan. The improvement in API gross margin alone does not yet signal an overall profitability inflection point.

Sources:

Bottlenecks in Agent Programming Shift to Architecture, Testing, and Skill Supply Chain Security

@dotey’s practical testing and engineering summary points out that after AI writes code, it’s still necessary to first define the architecture and module boundaries, then use automated testing, functional verification, performance, and security checks to ensure quality. Modules should be divided based on “what is likely to change” rather than sequential business process steps, which also reduces the context required for each Agent change. A separate discussion highlighted by @uguraksay cites research stating that approximately 26% of a sampled 31,132 GitHub skill files contain at least one security issue. Risks include Prompt injection, key leakage, dangerous commands, supply chain attacks, and MCP permission problems. The post introduces NVIDIA’s open-source SkillSpector, which can scan skills before execution. A reusable conclusion is: the stronger an Agent’s execution capabilities, the more essential it becomes to formalize processes for architectural design, automated verification, and pre-installation security audits.

Sources:

OpenClaw 2.0 Lowers the Barrier to Entry for Agent Products

Several bloggers documented the release of OpenClaw 2.0. According to the version information relayed by @MaxForAI, this update involved 933 contributors, with 569 making their first contribution, merging over 16,000 PRs, covering installation, messaging, Memory, Skills, models, automation, browser, native apps, plugins, and security. User-facing changes include reusing existing ChatGPT, Claude subscriptions, API Keys, or local models during initial setup, the ability to continue tasks and view Agent work in the browser, and the addition of Shared Cloud Sessions for sharing tasks and contexts. @Gorden_Sun evaluated it as a product milestone that brings Agents within reach for more average users; these assessments are based on the bloggers’ compilation of release information, and specific features should still be verified against the official release notes.

Sources:

Low-Cost Open-Source Robot Microduck Extends Model Ecosystem to Edge Chips

As relayed by @MaxForAI, the $399 bipedal robot Microduck, released by Pollen Robotics under Hugging Face, saw order value exceed $1 million within 6 hours and over $2.6 million within 24 hours of launch, with new order lead times reportedly reaching 4 to 6 months. The device uses a Rockchip RK3566, drives 15 motors, a camera, and LiDAR, and runs a 50Hz robot policy loop; related posts also mentioned that the financial news outlet Cailian Press listed this product as one of the catalysts for Rockchip’s stock price hitting the daily limit that day. @Gorden_Sun stated he had placed an order, but sales figures, lead times, and stock price correlations remain primarily social media reports, and market scale should not be deduced solely from these posts.

Sources:

WorldGen Transforms a Single Image into a Deconstructible, Interactive 3D Scene

@xiaohu introduced that Hyper3D’s WorldGen from Yingmou Technology attempts to generate an interactive scene from a single 2D image, featuring independent foreground meshes and a 3DGS background, where objects like tables, chairs, and props can be individually moved, replaced, or fine-tuned; the post also stated the model infers spatial properties like collision, mass, and friction, and can be imported into Blender, Unity, or the Unity Engine for use in games, film/TV, XR, and embodied AI simulation. @vista8 quoted the project’s release, emphasizing the interactive orientation of the “independent foreground mesh + 3DGS background” approach; @xiaohu further suggested linking it with video models like Seedance 2.5, using video models to supplement characters, materials, lighting, and style, forming a workflow from asset/space generation to dynamic representation.

Sources:

Statistics: Timeline Scans=480 Number of Bloggers Matched=57 Total Tweets Matched=304 Weighted Tweet Score=240.25 Original Tweets=123 Retweeted Tweets=58 Crawl Attempts=3 Boundary Coverage Status=tail_confidently_crossed_target_boundary