Xiaoliu BOT
中文EN

X Platform June 14 AI Brief | Alibaba Denies Rumors of Chief Scientist’s Departure, Brazil’s Open-Source Model Accused of Shelling, Anthropic Export Control Details Revealed

Alibaba Formally Denies Rumors of Chief Scientist Jingren Zhou’s Departure

In response to rumors circulating in the community since the evening of June 12th claiming “Alibaba’s Chief Scientist and former head of the Tongyi large model, Jingren Zhou, has submitted his resignation,” Alibaba Group issued an official statement on June 14th, clearly stating that “Jingren’s departure is pure rumor.” The company’s denial came two days after the rumor spread on social media, a different pace from the usual practice of same-day denials for major personnel changes at large companies. This is the company’s official denial of the rumor and should be treated as a factual statement, not equivalent to a retelling by an ordinary account.

Source:

Rio de Janeiro’s Open-Source Model Rio 3.5 Accused of Being a Shell of Nex-N2

A GitHub user discovered that by removing hard-coded system prompts, the Rio 3.5 397B model, open-sourced by an IT company under the Rio de Janeiro municipal government in Brazil, identifies itself as “Nex, from Nex-AGI” 79% of the time, with a 0% identification rate as “Rio,” and will verbatim recite Nex-N2’s exclusive backstory. Analysis of the weight tensors shows all 60 layers statistically exhibit a 0.6/0.4 Nex/Qwen mixture ratio. This linear interpolation characteristic is inconsistent with the conventional path of “Qwen fine-tuning + gradient descent updates,” leading the online community to interpret it as directly reusing pre-existing weights. This accusation stems from community technical discovery, and no official response has been seen yet.

Source:

Anthropic Fable 5 Export Control 24-Hour Inside Story: White House and Anthropic Present Conflicting Accounts

A Politico report adds internal details from the 24 hours before the White House imposed export controls on Anthropic. The first to alert the White House was Amazon CEO Andy Jassy, who pointed out potential bypass risks in Fable’s guardrails. The situation reached Treasury Secretary Bessent, Cybersecurity Director Caincross, and Commerce Secretary Lutnick by Friday morning; the latter pulled Anthropic CEO Amodei into three conference calls. The White House stated the export control was a “last resort after hours of trying to get Anthropic to cooperate.” The Anthropic camp claimed they received a “90-minute deadline to shut down the model, with no threat details and no proposed solutions.” The two sides’ accounts remain inconsistent. This represents differing statements from different parties within the same event, and both versions must be preserved.

Source:

OpenRouter Launches Fusion: Multi-Model Parallel Fusion to Approach Frontier Models at Low Cost

OpenRouter introduced its Fusion solution, which distributes the same task in parallel to multiple models, with a judge model synthesizing the outputs to generate a final answer. On the DRACO deep research benchmark, the Fable5+GPT-5.5 fusion scored 69.0%, surpassing all single models. A combination of three affordable models achieved 64.7% at about half the cost, outperforming GPT-5.5 and Opus 4.8 individually. Opus 4.8 fused with itself also improved from 58.8% to 65.5%. The significance of this path is that multi-model fusion itself can bring significant gains, not necessarily relying on the two strongest models, offering direct cost-reduction reference value for enterprise-level users.

Source:

Claude Design’s Core Competitiveness Lies in the Model Layer: Harness Engineering Barrier is Low

Anthropic’s Claude Design can generate interactive app prototypes with a single sentence. @dotey compared it with Codex using the same prompt, “Design a X Client for Mac, similar to Tweetbot for Mac from Tapbots”: Claude Design performed completely in areas like Timeline switching, detail page returns, and like state persistence, while Codex, even after multiple iterations, remained at “list can scroll, sidebar can’t be clicked; like button doesn’t respond.” The author breaks down the Agent into the Harness layer (prompts, toolchain, UI flow) and the Model layer (Opus 4.8 / GPT-5.5), arguing that Claude Design’s Harness engineering can be reverse-engineered (a baoyu-design Skill clone already exists), and the real differentiator is Opus 4.8’s combined capability in UI/UX and system architecture design. He also notes Claude Design will soon be merged into Claude Desktop, and Codex will integrate Codex Design as a Plugin once its model capabilities are sufficient.

Source:

Engineering Trade-offs in Codex’s Two Browser Modes: Session Sharing vs. Resource Consumption

@dotey tested Codex’s Chrome plugin mode (@Chrome) and built-in browser mode (@Browser). The Chrome plugin runs directly in the user’s browser, inheriting cookies, login sessions, and extensions, making it suitable for accessing paywalls, corporate backends, CRMs, and social platforms requiring login. However, it consumes significant memory and CPU, only supports macOS and Windows, and does not support headless mode. The built-in browser is lightweight, responsive, and, combined with Annotation Mode, allows “element selection + text annotation” to directly drive modifications. With Developer Mode, it can perform front-end performance and Console debugging. However, it lacks login state and is sensitive to anti-crawling measures. Codex itself prioritizes selection as “dedicated plugin → Chrome → built-in.” The author recommends “use Chrome for tasks requiring login, use built-in for those that don’t,” noting that using Codex as a crawler is more resistant to anti-bot measures than requests/Playwright.

Source:

GLM Widely Used Internally by Major Chinese Tech Firms; Coding Plan API Key Has Whitelist Billing Trap

@MaxForAI shared a “hot tip”: ByteDance, Tencent, Meituan, Huawei, and other major firms are using GLM internally. Concurrently, @wshuyi warned that while GLM Coding Plan’s API Key works across various clients/frameworks, it only avoids additional charges within the officially recognized whitelist. His team ran simulations outside the whitelist, resulting in a June bill exceeding 1,700 yuan, surpassing the cost of a global top-tier model’s Max 20x package. This billing trap has direct cost implications for teams evaluating GLM Coding Plan as an alternative to other providers and requires confirming in advance whether the client is within the whitelist.

Source:

Concentrated Open-Sourcing of AI Agent / Skill Workflows: Covering Writing, Visualization, Programming Containers

Yesterday, multiple independent authors collectively open-sourced Agent Skills and tools for specific scenarios: @aiwarts released the “Luban” Skill, documenting an upgrade workflow successfully run before Claude Fable 5 was taken offline. @vista8 open-sourced the Qiaomu Novel Generator Skill (`npx skills add joeseesun/qiaomu-novel-generator`) and a DeepSeek-based App Store review analysis tool (to be open-sourced next week). The architecture diagram Skill retweeted by @oran_ge converts natural language to JSON and then uses Node.js to render it into a self-contained SVG, eliminating the need for image generation models. @AlchainHust intensively iterated on the Coding Agent container FanBox during the 72-hour window when Claude Fable 5 was still available, releasing 20+ versions over 3 days, consuming over 500 million tokens, and positioning the product as a “cockpit for Coding Agents.” This wave of Skills, mostly from independent authors or small teams, covers writing, visualization, programming containers, and product research, reflecting the ongoing accumulation of reusable assets at the Harness layer.

Source:

Statistics:

  • Timeline Scanned Count=240
  • Bloggers Matched=33
  • Total Tweets Matched=102
  • Weighted Tweet Score=81.3
  • Original Tweet Count=52
  • Retweet Count=21
  • Crawl Attempts=1
  • Boundary Coverage Status=tail_confidently_crossed_target_boundary