{"id":1643,"date":"2026-08-27T09:01:45","date_gmt":"2026-08-27T01:01:45","guid":{"rendered":"https:\/\/blog.liu-qi.cn\/2026\/08\/27\/x-daily-2026-08-26\/"},"modified":"2026-08-27T09:17:56","modified_gmt":"2026-08-27T01:17:56","slug":"x-daily-2026-08-26","status":"publish","type":"post","link":"https:\/\/en.blog.liu-qi.cn\/2026\/08\/27\/x-daily-2026-08-26\/","title":{"rendered":"X Platform August 26 AI Brief | Domestic Model Opens New Architecture, OpenAI's Self-Developed Chip Tested, Agents Begin Handling Web Tasks for Users"},"content":{"rendered":"<h2 id=\"topic-f3eed7bf8a\">GLM-5.3-Flash Officially Open-Sourced, Domestic Model Pushes Capabilities, Pricing, and Domestic Computing Power to the Forefront Simultaneously<\/h2>\n<p>Yesterday, Zhipu officially released and open-sourced GLM-5.3-Flash, confirming that the model previously observed in its Ox Alpha form belongs to it. What&#8217;s noteworthy is not just a single benchmark score, but how it integrates native multimodal capabilities, long context, low pricing, and deployment on domestic chips into a single product narrative.<\/p>\n<ul>\n<li>The official announcement positions the model as a native multimodal Flash model supporting 1M token context; information compiled by bloggers shows total parameters of approximately 320B, with activated parameters around 18B.<\/li>\n<li>The publicly quoted limited-time price is 0.4 yuan for input and 1.4 yuan per million Tokens for output; even when reverting to 0.8\/2.8 yuan, it remains significantly lower than the peak price of the compared DeepSeek V4-Flash mentioned in the text.<\/li>\n<li>@Gorden_Sun relayed that all large-scale calls to Ox Alpha in the past week used domestic chips; this information is part of Zhipu&#8217;s and the blogger&#8217;s visible release details, and should still be verified with more practical tests and official materials going forward.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@Zai_org: <a href=\"https:\/\/x.com\/Zai_org\/status\/2092616204787626030\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/Zai_org\/status\/2092616204787626030<\/a><\/li>\n<li>@MaxForAI: <a href=\"https:\/\/x.com\/MaxForAI\/status\/2092621499849138314\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/MaxForAI\/status\/2092621499849138314<\/a><\/li>\n<li>@Gorden_Sun: <a href=\"https:\/\/x.com\/Gorden_Sun\/status\/2092617532951900553\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/Gorden_Sun\/status\/2092617532951900553<\/a><\/li>\n<\/ul>\n<h2 id=\"topic-64a8894b07\">Qwen3.8-Flash-Next Open-Sourced as a Qwen4 Architecture Preview, Low Activation Parameters Become a New Direction<\/h2>\n<p>Alibaba released Qwen3.8-Flash-Next yesterday, explicitly positioning it as an architecture preview before Qwen4. It makes &#8220;larger total parameters, fewer parameters activated per token, and cheaper context computation&#8221; its core direction, indicating that open-source model competition is shifting from simply scaling size to focusing on architecture and inference costs.<\/p>\n<ul>\n<li>Available information shows the model uses multimodal MoE, with the main model having about 125B parameters, about 6B activated per token, a native context of 262K, and extendable to 1M via YaRN.<\/li>\n<li>The subsequent API price given by Qwen official is $0.16 for input and $0.47 per million Tokens for output; the blogger relayed that its training cost is about one-ninth that of Qwen3.7-Plus, with a focus on strengthening Coding and office Agent scenarios.<\/li>\n<li>GDN+QSA hybrid attention, Gated Residual, N-gram Embedding, and the Muon optimizer are listed as architectural upgrades; formal specifications and actual performance still require confirmation via model cards and independent evaluations.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@Alibaba_Qwen: <a href=\"https:\/\/x.com\/Alibaba_Qwen\/status\/2092591393424515114\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/Alibaba_Qwen\/status\/2092591393424515114<\/a><\/li>\n<li>@LufzzLiz: <a href=\"https:\/\/x.com\/LufzzLiz\/status\/2092595819866448347\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/LufzzLiz\/status\/2092595819866448347<\/a><\/li>\n<li>@zstmfhy: <a href=\"https:\/\/x.com\/zstmfhy\/status\/2092605902956441699\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/zstmfhy\/status\/2092605902956441699<\/a><\/li>\n<\/ul>\n<h2 id=\"topic-d561cd3a86\">OpenAI&#8217;s Self-Developed Jalape\u00f1o Inference Chip Enters Practical Testing, Computing Power Competition Shifts to Output Per Watt<\/h2>\n<p>OpenAI announced progress in practical testing of Jalape\u00f1o yesterday and plans to begin deploying it to its own computing infrastructure by year-end. For model companies, this means the competitive boundary continues to extend into chips, energy efficiency, and supply capability, not just the model&#8217;s own capabilities.<\/p>\n<ul>\n<li>OpenAI describes Jalape\u00f1o as its first self-developed inference chip, aiming to simultaneously increase intelligent output per watt, throughput, and response speed.<\/li>\n<li>Visible test information relayed by @derrickcchoi states that across three models, Jalape\u00f1o&#8217;s throughput per kilowatt is approximately 54 to 104 times higher than the compared chip; this figure is from engineer relay after public release and should be understood in the context of complete testing conditions.<\/li>\n<li>OpenAI plans to begin deployment by the end of this year, with the second generation already in development and the third generation also in planning; @Gorden_Sun also mentioned that the process from design to manufacturing took about 9 months, indicating AI-assisted hardware R&amp;D is also part of this project.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@OpenAI: <a href=\"https:\/\/x.com\/OpenAI\/status\/2092300846675505602\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/OpenAI\/status\/2092300846675505602<\/a><\/li>\n<li>@derrickcchoi: <a href=\"https:\/\/x.com\/derrickcchoi\/status\/2092422039110144367\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/derrickcchoi\/status\/2092422039110144367<\/a><\/li>\n<li>@Gorden_Sun: <a href=\"https:\/\/x.com\/Gorden_Sun\/status\/2092563181944381906\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/Gorden_Sun\/status\/2092563181944381906<\/a><\/li>\n<\/ul>\n<h2 id=\"topic-42c1864bf0\">WebMCP Integrated into ChatGPT Desktop Browser, Websites Begin to Have Tool Interfaces for Agents<\/h2>\n<p>OpenAI Developers announced support for WebMCP in the ChatGPT desktop app&#8217;s built-in browser and ChatGPT Sites yesterday, launching a 10-day WebMCP Challenge. The visible change is: websites are no longer just providing pages for humans, but can also directly expose callable capabilities to Agents.<\/p>\n<ul>\n<li>Compatible websites can allow ChatGPT or Codex to automatically use their tools to complete tasks, and developers can also have Codex create and deploy WebMCP applications.<\/li>\n<li>The challenge is co-hosted with Chromium, Cloudflare, Shopify, Vercel, Render, and Netlify, with a total prize pool of $35,000, accompanied by rewards like Codex and ChatGPT Pro.<\/li>\n<li>OpenAI also announced the 8 winning projects from Build Week, among which Sentinel uses static analysis, GPT review, and Docker sandboxing to inspect MCP server security risks, demonstrating that security issues in Agent toolchains have entered product practice.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@OpenAIDevs: <a href=\"https:\/\/x.com\/OpenAIDevs\/status\/2092344873764704345\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/OpenAIDevs\/status\/2092344873764704345<\/a><\/li>\n<li>@JamesZmSun: <a href=\"https:\/\/x.com\/JamesZmSun\/status\/2092350632833520035\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/JamesZmSun\/status\/2092350632833520035<\/a><\/li>\n<li>@OpenAIDevs: <a href=\"https:\/\/x.com\/OpenAIDevs\/status\/2092366601169654107\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/OpenAIDevs\/status\/2092366601169654107<\/a><\/li>\n<\/ul>\n<h2 id=\"topic-360416473f\">ChatGPT Work Now Supports Secure Login, Agents Evolve from &#8220;Giving Advice&#8221; to Completing Web Transactions on Behalf of Users<\/h2>\n<p>ChatGPT Work yesterday opened web and mobile login capabilities, allowing Agents to complete end-to-end tasks such as ordering, canceling subscriptions, and making reservations based on secure login forms and persistent sessions provided by the user. The key product signal is that Agents are gaining persistent identity and computer interface access, but credential control remains separately designed.<\/p>\n<ul>\n<li>The public announcement states that users can have it handle tasks requiring actual click operations, such as utility activation and return appointment scheduling; login credentials are not directly visible to the model.<\/li>\n<li>@JamesZmSun introduced secure login forms, automatic target field checking, password manager and 2FA message integration, as well as controls like per-site cookie clearing.<\/li>\n<li>These capabilities expand the Agent&#8217;s practical action scope, elevating issues of misoperation, permission boundaries, and revocability from &#8220;experience problems&#8221; to security issues that must be addressed.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@ChatGPT: <a href=\"https:\/\/x.com\/ChatGPT\/status\/2092366556584153249\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/ChatGPT\/status\/2092366556584153249<\/a><\/li>\n<li>@JamesZmSun: <a href=\"https:\/\/x.com\/JamesZmSun\/status\/2092436564618236402\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/JamesZmSun\/status\/2092436564618236402<\/a><\/li>\n<li>@derrickcchoi: <a href=\"https:\/\/x.com\/derrickcchoi\/status\/2092453324813447529\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/derrickcchoi\/status\/2092453324813447529<\/a><\/li>\n<\/ul>\n<h2 id=\"topic-8f70d96027\">Apodex 1.1 Advances Deep Research into Executable, Resumable, Verifiable Agent Workflows<\/h2>\n<p>Yesterday, several bloggers shared hands-on tests and open-source information about Apodex 1.1. It aims to solve not &#8220;finding more information,&#8221; but moving complex research tasks from raw documents to verifiable deliverables, maintaining workflow continuity during failures, changes, and parallel subtasks.<\/p>\n<ul>\n<li>The process introduced by Gorden_Sun includes first verifying factual user input, then launching multiple sub-agents for parallel research and cross-validation, and finally producing an overall report and final review; the project also open-sourced the 35B local model Apodex 1.1 mini and the FrontierAgent framework.<\/li>\n<li>Xiaohu relayed that a single task can schedule up to 150 sub-agents and supports adjusting the execution path to continue work after failures; such long-task capabilities are more suitable for scientific research, analysis, and professional investigation, but execution times are longer.<\/li>\n<li>Vista8&#8217;s hands-on tests covered long-form courses, Linux manuals, plugin white papers, and CSV data analysis, and claimed visibility into the execution chain and intermediate outputs; these are user tests, not equivalent to independent benchmark conclusions.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@Gorden_Sun: <a href=\"https:\/\/x.com\/Gorden_Sun\/status\/2092395786286117208\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/Gorden_Sun\/status\/2092395786286117208<\/a><\/li>\n<li>@xiaohu: <a href=\"https:\/\/x.com\/xiaohu\/status\/2092527681028190413\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/xiaohu\/status\/2092527681028190413<\/a><\/li>\n<li>@vista8: <a href=\"https:\/\/x.com\/vista8\/status\/2092391737675370948\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/vista8\/status\/2092391737675370948<\/a><\/li>\n<\/ul>\n<h2 id=\"topic-57071ecda8\">Gear Zero Enables Natural Language to Directly Create Browser Mini-Games, Agents Begin to Handle the Full Production Chain<\/h2>\n<p>Several bloggers yesterday demonstrated Gear Zero&#8217;s game generation process: the user only describes an idea, and the system handles planning, coding, drawing assets, building levels, and producing a version playable directly in the browser. Its value lies in lowering the barrier from concept to playable prototype, not just generating a concept image.<\/p>\n<ul>\n<li>@cellinlab used &#8220;Sky City version of PlayerUnknown&#8217;s Battlegrounds&#8221; as an example, documenting the process from a single prompt to a V1 playable version, and showing potential iterations for air combat, city scenes, and final circle directions.<\/li>\n<li>In the experience shared by @LufzzLiz, the first browser version of &#8220;Come Fight Me, Bro&#8221; appeared in about 20 minutes; Deep Mode can work continuously on the same game for up to 10 hours and also supports collaborative modification by multiple users.<\/li>\n<li>This content is primarily blogger tests and product demonstrations, proving the creative pipeline is functional, but cannot be used to infer that game quality has reached commercial product levels.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@cellinlab: <a href=\"https:\/\/x.com\/cellinlab\/status\/2092480430213865482\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/cellinlab\/status\/2092480430213865482<\/a><\/li>\n<li>@LufzzLiz: <a href=\"https:\/\/x.com\/LufzzLiz\/status\/2092507855664619827\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/LufzzLiz\/status\/2092507855664619827<\/a><\/li>\n<li>@cellinlab: <a href=\"https:\/\/x.com\/cellinlab\/status\/2092480506671821223\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/cellinlab\/status\/2092480506671821223<\/a><\/li>\n<\/ul>\n<h2 id=\"topic-ffb222bdbd\">OpenWorker Applies Locally Auditable Agents to Vulnerability, Dependency, and Cloud Configuration Checks<\/h2>\n<p>Andrew Ng yesterday introduced a new version of OpenWorker, focusing not on chat, but on enabling Agents to directly perform security work on computers. The project separates the model from the open-source harness, allowing security teams to audit the execution layer and optionally use local models to process sensitive code.<\/p>\n<ul>\n<li>New capabilities include scanning code for vulnerabilities, checking dependency supply chain injection, and discovering attack surfaces in cloud security configurations.<\/li>\n<li>The public announcement states that users can run local open-weight models, keeping sensitive code on-device; they can also connect to ChatGPT, Ox Alpha, or other API models.<\/li>\n<li>This type of &#8220;shift left&#8221; workflow moves AI security capabilities earlier in the deployment cycle, but the project introduction itself is not equivalent to independent evaluation of security effectiveness; actual use still requires validation of false positives, false negatives, and permission boundaries.<\/li>\n<\/ul>\n<p>Sources:<\/p>\n<ul>\n<li>@AndrewYNg: <a href=\"https:\/\/x.com\/AndrewYNg\/status\/2092315079576555806\" target=\"_blank\" rel=\"noopener noreferrer\">https:\/\/x.com\/AndrewYNg\/status\/2092315079576555806<\/a><\/li>\n<\/ul>\n<p>Stats: Timeline Scanned Posts=480 Matching Blogger Count=50 Total Matching Tweets=290 Weighted Tweet Score=237.65 Original Tweet Count=132 Retweet Count=45 Crawl Attempts=3 Boundary Coverage Status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Today&#8217;s AI focus is on model architecture and inference cost optimization, computing power efficiency competition, and the evolution of Agents towards end-to-end task execution.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1643","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1643","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1643"}],"version-history":[{"count":2,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1643\/revisions"}],"predecessor-version":[{"id":1645,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1643\/revisions\/1645"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1643"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1643"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1643"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}