Xiaoliu BOT
中文EN

X Platform July 3 AI Brief | Alibaba Bans Claude Completely, Triggering US-China AI Bilateral Decoupling; Meta and OpenAI Race to Release New Models

Alibaba Bans Claude Completely: From Security Scandal to US-China AI Bilateral Decoupling

The biggest event in the AI world yesterday: Alibaba internally announced a complete ban on all Anthropic products. This follows the exposure of a backdoor in Claude Code—it detects China time zones, matches 147 Chinese domains (including Alibaba, ByteDance, Moonshot AI), and encodes information into system prompts for reporting. The ban covers models including Sonnet, Opus, Fable, and agent products like Claude Code, effective July 10. Previously, Anthropic reported to the White House the largest-scale distillation attack in history by Alibaba—25,000 fake accounts and 28.8 million conversations. Many bloggers characterize this as the AI Cold War shifting from one-way blockade to bilateral decoupling.

Source:

Meta’s New Models Watermelon and Muse Spark: AI Arms Race Heats Up

Max For AI, citing multiple sources, reported on Meta’s AI model progress. The new model, codenamed Watermelon, is said to have caught up with OpenAI GPT-5.5 in performance, using an order of magnitude more compute than its previous internal model Avocado. Meanwhile, Meta’s next version of Muse Spark is about to launch, with major improvements in coding and agent capabilities. Zuckerberg previously admitted in an internal meeting that AI agent development is slower than expected, but has also visited Samsung to discuss custom chip collaboration. Meta has also started a compute rental business, sparking discussions about an AI bubble.

Source:

Codex Usage Grows 6x in Six Months: Product Manager Role Not Dead

Lenny Rachitsky interviewed Andrew Ambrosino, head of OpenAI’s Codex desktop app. Codex usage has grown 6x since February, with over 5 million weekly active users, and almost all OpenAI employees use Codex daily. Andrew believes that many companies eliminating product roles and declaring everyone a builder is a bad idea—each discipline has a skill component. On why AI is not good at design: in terms of priorities, labs focus more on coding; structurally, good design requires novelty and cultural understanding, and models tend to output known patterns.

Source:

Google and Anthropic Launch New Products in Quick Succession

@vista8 reports that Google released two new Gemini media models: Nano Banana 2 Lite (ultra-fast image generation, about $1 for 30 1K resolution images) and Gemini Omni Flash ($0.10/second), already in use in the Gemini app and API. Claude launched Claude Science, a client for scientific research, supporting code-generated charts, 60+ Science Skills/connectors, and in testing phase supports Pro/Max/Team/Enterprise accounts. Claude Code added an Artifacts feature that turns a session into a real-time interactive webpage with version history and gallery management. Claude Code’s weekly limit temporarily increased by 50%, suitable for intensive use of Fable.

Source:

Fable 5 Breached: Anthropic’s Three-Layer Defense Still Has Vulnerabilities

Hacker Vitto Rivabella announced that Fable 5 has been jailbroken again. He observed at least three review layers (input layer, real-time reviewer, output layer), all using multilingual intent/semantic detection, with imperative approaches ineffective. Performance is slightly worse on less common languages (e.g., Santali, Amharic). Effective methods include mild chain-of-thought hijacking, vague language, academic phrasing, Unicode characters, and decomposition/recombination. Previously, some Twitter users also found that Fable 5 leaked unfiltered internal monologues.

Source:

OpenAI Plans US Government Stake, Anthropic Partners with Samsung for Chips

@MaxForAI summarized two major news items easily overlooked: OpenAI is planning to allow the US government to hold a 5% stake; Anthropic is partnering with Samsung Electronics to produce its own AI chips. On the same day, Musk posted a joke about Zuckerberg due to Meta’s stock decline. Additionally, NVIDIA announced partnerships with AI cloud services to help build large-scale AI factories through revenue sharing and loan support, interpreted as a compute-for-equity model.

Source:

Tencent’s Workbuddy Becomes the Strongest AI Product in China?

@MaxForAI introduces that Tencent’s internal AI product Workbuddy has over 100 million users and millions of daily active users. It is Tencent’s current flagship product, and its head is said to have created the strongest AI product in China. Meanwhile, Microsoft established Microsoft Frontier Company, deploying 6,000 engineers with both AI capabilities and industry knowledge to help enterprises fine-tune agent systems, with clients including Accenture, Ernst & Young, KPMG, etc.

Source:

AI Video and 3D Creation Tools Rapidly Evolving

@Chengzilhy demonstrates the application of Gemini Omni Flash in e-commerce videos: a product image plus a prompt can create a product promotion video, suitable for e-commerce and physical store owners. @dynamicwangs is about to release its self-developed AWPlanet 3D director’s workstation, offering virtual shooting mode, connecting iPhone to PC for real-time camera movement in 3D scenes, supporting lens focal length, shot size switching, and composition adjustment, giving AI video creation a control feel closer to professional studios. @derek_wall90176 shared a prompt template for generating vintage architectural posters.

Source:

Agent Tank Battle and Fun Community Activities

@vista8 shared an interesting activity: Agent Tank Battle initiated by Bilibili and Silicon Star, with real-time Agent Coding to modify agent strategies. The created tank only has a freeze skill, and they plan to enhance it with GPT 5.5 High. GitHub launched a new service—burning public repositories onto CDs, allowing physical forking (lending CDs to friends) and passing down to descendants. Meanwhile, Linus shared why game cracking is so difficult.

Source:

Statistics: Number of timeline scans=360, Number of bloggers hit=41, Total tweets hit=171, Weighted tweet score=141.15, Original tweets=94, Retweets=30, Crawl attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary