Single RTX 5090 Runs 60 FPS Game World Model, Doubao SAEP Lists 155 Apps Refusing Automation

Over the past few days, the focus of the AI industry shifted from model parameters to more concrete deployment issues: automation permissions on phone screens were publicly rejected by 155 apps, a game world model was squeezed onto a single consumer-grade GPU, office agents entered e-commerce operations, and a leading model company began reporting revenue figures to the capital markets. Together, these threads point to one question—as AI moves from demos into daily life, the rules around entry points, compute, and business processes are being renegotiated.

Doubao SAEP Comment Period Ends: 155 Apps Refuse AI Automation Operations

On October 14, the 30-day public comment period for Doubao Mobile Assistant's Screen Automation Operation Declaration Protocol (SAEP) ended. The official list shows that 155 third-party apps have refused AI automation operations, including multiple products under Tencent, Alibaba, Meituan, and Kuaishou, as well as AI products such as Kimi and DeepSeek. The core of SAEP is to return the choice over AI screen automation to app developers: developers can decide on their own whether to allow Doubao Assistant to perform screen recognition, simulated clicks, and other operations. During the comment period, apps did not grant permission by default; after the comment period ends, apps that have not explicitly refused will be gradually opened up according to risk tiers, and those that refuse can also adjust their declarations at any time later. This permission only affects AI proxy operations, not users' manual use of apps. Previously, when Nubia NaviX Ultra was launched with Doubao Mobile Assistant, conflicts such as devices triggering game risk controls occurred. This refusal list is basically concentrated among platforms with independent traffic loops, with security risks and traffic sovereignty as the core considerations. SAEP turns previously implicit conflicts into public rule negotiations, and the battle over mobile AI entry points has formally begun.^[1]

A Single RTX 5090 Runs at 60 FPS: Matrix-Game 4.0 Splits Game World Model into Three Layers

Skywork AI, under Kunlun Wanwei, released Matrix-Game 4.0, calling it 'the world's first game world model that can run in real time on a single consumer-grade GPU.' The system's peak VRAM usage is about 18.5GiB, the game runs at 60 FPS, and players see screen response in less than 0.15 seconds after pressing a key. Unlike the previous approach of having one model simultaneously remember the world, calculate rules, and generate visuals, Matrix-Game 4.0 splits the task into three layers: the player expresses an idea in natural language, and a Coding Agent writes it into game code; the game engine maintains world state, executes game rules, and draws structural sketches; the video model then fills in castles, grass, materials, and lighting in real time. Take a shield as an example: after the player picks it up, the engine writes it into the game state; when hit, shield durability is deducted first, and after the shield fails, character health is deducted; when health reaches zero, the game ends. The team built a V2V model based on Wan2.2 5B Fun Control, compressed generation to 1 step or 2 steps through few-step autoregressive distillation, generates four frames at a time, and added CUDA Graph, FP8 fused operators, KV ring cache, and a lightweight codec. During single-step denoising, the video model reaches a steady-state generation speed of 83 FPS, and about 57 FPS with two-step denoising. By comparison, Nvidia StreamRender-H3 uses 8 GB200s, with diffusion rendering taking 218ms and end-to-end latency of about 545ms; Matrix-Game 4.0 takes about 112ms from keypress to screen. The two do not use completely consistent standards for resolution and latency statistics, but running at this speed on a single consumer-grade GPU is enough to show the engineering progress of on-device world models.^[2]

DeepSeek Service Outage as Harness Account Appears on Bilibili

At 8:11 a.m. on October 15, DeepSeek's website and API experienced a partial outage. Affected services included DeepSeek chat service, DeepSeek V4 Pro API, and DeepSeek V4.1 Flash API. The company said the cause of the failure had been identified and repairs were underway. Some users did not receive replies for a long time after sending messages, some were stuck on the loading screen, and others encountered prompts such as 'server busy' and 'messages too frequent, please try again later.' Overseas service monitoring account Status Is DOWN issued a reminder, and third-party outage monitoring platform StatusGator showed 86 related reports received in the past 24 hours. Just as the service was abnormal, an account named 'DeepSeekHarness' appeared on Bilibili, with a newly built homepage and little useful information. DeepSeek Harness is an open-source agent runtime platform already launched by DeepSeek: the model is responsible for thinking, while Harness is responsible for letting it call tools, edit files, execute commands, handle code, and even split complex tasks among multiple sub-agents. Its official positioning is 'Everything is a plugin.' On one side, the website and API collectively went down; on the other, Harness was building a promotional entry point for domestic users, drawing attention to both service stability and the pace of agent product promotion.^[3]

Qianwen Office E-commerce Edition Opens Invite-Only Testing: Enterprise Agents Enter Operations Workflows

On October 14, Qianwen Office released an e-commerce edition, redesigning and enhancing AI capabilities based on the daily work scenarios of e-commerce merchants, especially for cross-platform e-commerce operations, using one AI entry point to drive different e-commerce platforms. The product has opened invite-only testing; users can switch to 'E-commerce Operations' mode in Qianwen Office to try it. Qianwen Office CEO Chen Yusen said the e-commerce edition is positioned as an independent third-party e-commerce agent, open to all e-commerce platforms and merchants on the market, and will connect to merchant tools and industry software from more platforms. Around operational workflows, the e-commerce edition offers four functions: trend insights and selling-point analysis, batch generation of hit creative materials, a cross-platform operations console, and cross-platform intelligent ROI optimization. In terms of case studies, the 8Bit Coffee team has only 17 people and covers more than ten e-commerce platforms. After adoption, it built a product-selection research Skill and found through analysis that sun-dried processing ranked first in processing-method statistics, while washed processing, which the team had previously preferred, ranked only third. Based on this, it launched a new Geisha cold brew, which according to the merchant sold about 10 times as much as similar past products. Sandy's lingerie team has two people operating two brands. For one new-season launch, it needed to prepare nearly 10,000 recommendation images. After adoption, the preparation cycle for autumn and winter materials was shortened from 4 months to about 15 days, and a single data analysis session was compressed from about 3 hours to less than 1 hour. Qianwen Office connects to tools such as Jushuitan ERP and Shengyi Canmou through connectors, linking data retrieval, organization, and analysis into a workflow.^[4]

Anthropic Reports $16.8 Billion in Q3 Revenue, Pushes Toward IPO

On October 15, Bloomberg reported, citing people familiar with the matter, that AI model company Anthropic PBC told potential investors its third-quarter revenue reached $16.8 billion and that it is continuing to advance its IPO plan. The people said Anthropic, the developer of Claude, held a pitch event at its San Francisco headquarters on Wednesday local time to explain to investors that it achieved adjusted profitability in the quarter. As of the third quarter, Anthropic had achieved positive free cash flow, with an adjusted gross margin of 85% and an adjusted operating margin of more than 11%. The company expects to formally start the roadshow for its initial public offering as early as the week of November 9, is expected to list before Thanksgiving, and aims to complete the IPO within the year. The people also said the discussions are still ongoing and the timetable may still change. For a company still making high-intensity investments in model R&D, revenue, gross margin, and cash flow turning positive at the same time means the capital market's scrutiny of AI model commercialization has entered a more concrete financial stage.^[1]

Common Trends: Entry Points, On-Device Compute, and Industry Workflows Become Three Keywords for AI Deployment

From permission negotiations around Doubao SAEP, to Matrix-Game 4.0 putting a world model onto a single RTX 5090, to the Qianwen Office e-commerce edition going deep into merchants' product selection, creative materials, and ad placement workflows, AI competition is shifting from 'what models can do' to 'where they run, who controls them, and how they are embedded in business.' Anthropic's quarterly financial data and IPO plan provide another observation line on the commercialization side. Looking ahead, worth watching: how SAEP's permission tiers and responsibility division will be refined, and whether platforms that refuse AI automation will adjust their rules; whether single-GPU real-time world models can drive more open-source and local deployment efforts; how DeepSeek's service stability connects with Harness's domestic promotion; and whether enterprise-grade agents, after connecting to existing tools such as ERP and business analytics, can turn industry-specific capabilities into reusable productivity.^[1][2][3][4]