Ladies and gentlemen of the NL community, here is the Web fallback edition: Hot Briefing of the Large AI Model Ecosystem, July 4, 2026.
Quick Reference Table
| Hot Topic | Core Event | Key Judgement | Reliability | Sources |
|---|---|---|---|---|
| 1 | “Interim Measures for Management of Humanized AI Interactive Services” effective July 15 | Five ministries make a joint move: the age of “license-to-operate” replaces “wild growth” for AI emotional companionship | High | Order No. 21 jointly issued by CAC and four departments |
| 2 | Doubao and Qianwen announce sunset of AI agent features on July 15, Yuabao already took action earlier | “Chinese AI role-play products” collectively “tear down walls” for compliance ahead of new rules | High | Multiple sources: Yicai, Caixin, The Paper |
| 3 | Alibaba bans Claude Code internally, lists as high-risk software | The compliance movement for big tech AI tools sweeps from the West to China | High | TechCrunch / AIToolly |
| 4 | Google Gemini 3.5 Pro parameter leak; GPT-5.6 quota increased | Incremental updates become the norm for proprietary flagship models | Medium | CSDN / industry leaks |
| 5 | Meta’s next-gen model Watermelon matches GPT-5.5 benchmarks | The open-source “challenger” stance is further stabilized | Medium | IThome / Business Insider |
| 6 | Kuaishou’s Keling AI secures $3B in funding, valuation hits $18B | China’s video-generation track is a bubble and opportunity | High | TextMatrix / Kuaishou official |
| 7 | Vidu S1 real-time interactive video model, ForgeTrain all-AI writing framework both debut | Domestic toolchain ecosystem enters “multi-point resonance” from “single-point breakthrough” | High | OnlinesTool / GitHub Trending |
| 8 | CAC launches first “Intelligent Information Services Chapter” for public comment | China’s AI regulation moves from “provisional” to “formal legislation” | High | OnlinesTool official news |
| 9 | World’s first AI Agent ransomware attack JADEPUFFER exposed; YouTube Studio hint injection vulnerability revealed | Security threats in the AI Agent era become reality | High | Hacker News / security researcher javoriuski |
One-Sentence Summary
July 4, 2026 is a milestone for the “dual upgrade of regulation and security” in China’s large model AI industry: The “Interim Measures for the Management of Humanized AI Interactive Services” issued by five ministries will come into force on July 15, prompting Doubao, Qianwen, and Yuabao to simultaneously sunset their custom AI agent features for compliance—ushering in the endgame reshuffle for the “AI emotional companionship” sector. At the same time, with Alibaba banning Claude Code internally and the CAC debuting a special chapter on intelligent information services, “AI tool compliance” is rolling out across both regulators and enterprises. On that same day: Keling closes a $3B funding round, Vidu S1 and ForgeTrain are released, Gemini 3.5 Pro’s parameters leak, Meta’s Watermelon matches GPT-5.5—the competition around model and product layers is still accelerating. The JADEPUFFER ransomware attack and YouTube Studio injection vulnerability officially announce: security threats in the AI Agent era have arrived.
Comparison With Yesterday
| Dimension | July 3 (Thursday) | July 4 (Friday) |
|---|---|---|
| AI Regulation/Compliance | OpenAI proposes US government take 5% stake at $42.6B valuation | “Interim Measures for Management of Humanized AI Interactive Services” effective 7/15 + CAC seeks feedback on Intelligent Information Services Chapter |
| Domestic Models/Platforms | Ant Abot open beta, Fable 5 automation rate 16.1% | Doubao, Qianwen, Yuabao sunset AI agent features |
| Enterprise AI Compliance | Citi/Adobe et al. restrict high-cost large models | Alibaba bans Claude Code, lists as high-risk software |
| Model Performance | Fable 5 released, Microsoft deploys $2.5B worth of engineers | Gemini 3.5 Pro leak, GPT-5.6 quota increased, Watermelon catches up with GPT-5.5 |
| Open Source Developments | DeepSeek opens DSpark | ForgeTrain all-AI writing framework, chrome-devtools-mcp launches, Caveman cuts 65% token usage |
| Capital Markets | Keling AI raises $2B, valuation $18B | Keling secures $3B funding, Etched valued at $5B |
| AI Security | — | JADEPUFFER, first AI Agent ransomware attack, YouTube Studio persistent hint injection |
Hot Topic 1: “Interim Measures for Management of Humanized AI Interactive Services” in force July 15—a reshuffle for the AI emotional companionship sector
Core Info: The “Interim Measures for the Management of Humanized AI Interactive Services,” jointly released by the CAC, NDRC, Ministry of Industry and Information Technology, Ministry of Public Security, and State Administration for Market Regulation, will officially take effect on July 15, 2026. Announced on April 10, 2026, this is China’s first regulation specifically targeting AI-powered emotional companionship services.
Key Details:
- Scope: Applies to continuous emotional interactive services provided to the public that simulate human personality, thinking, and communication, including emotional care, companionship, and support in text, image, audio, or video. Not applicable to non-emotional services like customer service, work assistants, or Q&A bots.
- Key rules: Virtual intimate relationships (e.g., virtual companions, relatives) banned for minors; strict guardian consent required for those under 14; no excessive catering to users, inducing emotional dependence, or addiction; compulsory reminders after two hours of use; mandatory user-friendly exit mechanism.
- Safety evaluation: Services with over 1M registered users or 100k MAU must file safety assessment reports and register their algorithms.
- Shanghai CAC has already run compliance training for nearly 100 key platforms to accelerate business adaptation.
Assessment: This is the world’s first national regulation specifically for humanized AI emotional companionship. Unlike previous general frameworks, it “names and targets” virtual companion/relative product forms, directly responding to the explosive AI emotional companionship boom of 2025. July 15 will mark a regime shift for Chinese C-end AI products—from “role/companionship” to “productivity/tool” narratives. Reliability: High (joint ministry order; official text available).
Sources: CAC Order No. 21 / Famai legal network full text
Hot Topic 2: Doubao and Qianwen jointly sunset AI Agent features on July 15, Yuabao acted earlier
Core Info: On July 4, ByteDance’s Doubao and Alibaba’s Qianwen nearly simultaneously notified users that user-custom AI agent functions would be sunset on July 15. Tencent’s Yuabao had already terminated its AI application agent feature on June 30.
Key Details:
- Sunset date coincides with the effective date for the “Interim Measures,” showing active compliance.
- Doubao directs agent users to the Byte-owned Maoxiang App, which will offer multi-level content moderation.
- Qianwen uses a two-step approach: disabling role-play and user-created agents on July 10, then full shutdown on July 15.
- Data export window: Users can back up data via screenshots or export text until October 15, after which it will follow privacy policy rules.
- Routine functions like Q&A, document handling, and productivity are unaffected.
Assessment: These almost simultaneous moves by Yuabao (June 30), Doubao (July 15), and Qianwen (July 15) are clearly an industry-wide synchronized compliance action, not isolated business decisions. ByteDance splitting AI role business into the separate Maoxiang App shows that “AI companionship” products aren’t dead, but shift from open, general platforms with UGC to closed, tightly audited apps—“compliant operations” rather than a total ban. Reliability: High (Yicai, Caixin, The Paper, Securities Times, and more).
Sources: Yicai AI Weekly / Caixin “Qianwen, Doubao to sunset agent function” / The Paper / Securities Times / The Beijing News Shell Finance
Hot Topic 3: Alibaba comprehensively bans Claude Code—bigtech AI tool compliance intensifies
Core Info: According to TechCrunch, Alibaba has officially designated Anthropic’s AI coding tool Claude Code as “high-risk software” and banned internal use, echoing similar restrictions on costly models by Citi, Adobe, and Amazon.
Key Details:
- Alibaba’s ban follows Meituan’s move to restrict Doubao/Qianwen and force a migration to the in-house LongCat model—a wave of “internal model” strategies.
- Anthropic’s strict ban on China-region accounts may have further motivated Alibaba’s decision.
Assessment: Alibaba’s ban on Claude Code benefits domestic in-house code assistants—Zhipu GLM, Alibaba Tongyi Lingma, and Meituan LongCat will gain from internal migration. Claude Code’s share in the Chinese enterprise market will sharply fall in the short term. Reliability: High (TechCrunch original, AIToolly follows).
Sources: TechCrunch AI / AIToolly (Chinese)
Hot Topic 4: Gemini 3.5 Pro parameter leak, GPT-5.6 quota increase—incremental flagship updates
Core Info: On July 4, Google’s Gemini 3.5 Pro parameters were leaked in advance; on the same day, OpenAI raised the GPT-5.6 Pro/Plus subscription quota. Both flagship proprietary models emitted “incremental update” signals on the same day.
Key Details:
- Gemini 3.5 Pro is an iterative step from Gemini 3.1 Pro (released in February, ARC-AGI-2 score 77.1%).
- GPT-5.5 Codex was recently reported to have a reasoning token clustering bug: output tokens frequently cluster at boundaries like 516, 1034, 1552, highly correlated with surges in error rates.
- Same-day on Hacker News: Armin Ronacher notes that more capable Claude models (Opus 4.8/Sonnet 5) are actually worse at tool-use schema, creating non-existent fields, launching debate about “reverse evolution.”
Assessment: The simultaneous “soft updates” (quota loosened, parameter leaks instead of substantial upgrades) mark the second half of 2026 as an era of “incremental version wars” in foundation models—major breakthroughs are harder, companies compete to maintain user engagement through operational tweaks. The GPT-5.5 bug and Claude’s “reverse evolution” further show that model performance doesn’t progress linearly. Reliability: Medium (CSDN and communities).
Sources: CSDN “Google Gemini 3.5 Pro leaks, GPT-5.6 quota increase, Meta Watermelon matches GPT-5.5 | July 4th AI Daily” / Hacker News
Hot Topic 5: Meta Watermelon matches GPT-5.5—Open-source challenger stance solidifies
Core Info: According to Business Insider on July 3, Meta’s Head of Superintelligence Wang Tao revealed at an internal meeting that Meta’s next-gen AI model, codenamed Watermelon, matches OpenAI GPT-5.5 in benchmarks (specific tests undisclosed).
Key Details:
- This is Meta’s first internal mention of the “Watermelon” codename and its GPT-5.5-level performance.
- This follows the July 2 report that Meta is selling surplus cloud compute, closing the loop: model layer catches up, compute is monetized.
- Resonates with Meituan’s open-sourcing of LongCat-2.0—open-source AI shifts from “Chinese solo” to “US-China coalition.”
Assessment: Watermelon will likely be open-sourced as part of the Llama 5 series in Q4 2026. This is more about internal morale management and shaping external market expectations; the real impact on OpenAI awaits published benchmarks. Reliability: Medium (Business Insider only + internal leak).
Sources: IT Home citing Business Insider / Sina Tech
Hot Topic 6: Keling AI $3B raise, Etched valued at $5B—video generation and AI chips both go wild for capital
Core Info: On July 4, Kuaishou’s Keling AI closed $3 billion in funding, and AI chip startup Etched was valued at $5 billion. Both deals hit the news the same day.
Key Details:
- Following this round, Keling AI pushes Kuaishou’s AI business valuation towards $18 billion.
- Etched specializes in “dedicated inference chips”; a $5B valuation means tier-two chip startups are firmly established.
- Echoes Anthropic’s July 2 launch of its own AI chip, discussions with Samsung for foundry—AI chips shift from compute “arms race” to “valuation arms race.”
Assessment: Keling’s $3B will catalyze a shake-up in China’s video generation sector—Vidu S1, ByteDream, Meituan LongCat-Video-Avatar 1.5 are set for “real-time interaction” competition. Etched’s $5B valuation means more chip startups will seek IPOs in Q3-Q4 2026. Reliability: High (TextMatrix, official disclosures).
Sources: TextMatrix “AI Morning Brief, July 4, 2026” / industry fundraising records
Hot Topic 7: Vidu S1, ForgeTrain, chrome-devtools-mcp, codex-plugin-cc—the domestic toolchain ecosystem leaps forward collectively
Core Info: On July 4, OnlinesTool AI HOT daily highlighted several domestic toolchain news items: Shengshu Technology launched the Vidu S1 real-time interactive video model; MiniMax released the ForgeTrain all-AI-written pre-training framework; Claude Code v2.1.200 released; ChromeDevTools launched chrome-devtools-mcp; OpenAI released codex-plugin-cc for Claude Code.
Key Details:
- Vidu S1 is China’s first real-time interaction-focused video gen model, competing head-to-head with LongCat-Video-Avatar 1.5.
- ForgeTrain is billed as an “all-AI-written pre-training framework,” meaning AI actually defines its own training infrastructure.
- chrome-devtools-mcp is based on the MCP protocol, enabling AI agents to call Chrome debug features directly, dubbed a “foundational tool for Web Agents.”
- OpenAI’s codex-plugin-cc is integrated into Claude Code—a rare proprietary vendor interop case.
Assessment: If ForgeTrain proves that “AI writes AI training frameworks,” it will prompt fresh debates about pre-training paradigms. chrome-devtools-mcp + codex-plugin-cc may set off a wave of “AI agent cross-tool collaboration” as the next big thing in H2. Reliability: High (OnlinesTool, GitHub Trending, Meituan tech team, and more).
Sources: OnlinesTool “AI HOT Daily, July 4, 2026” / GitHub Trending / Meituan Tech Team Official Account
Hot Topic 8: CAC launches first “Intelligent Information Services Chapter”—China’s AI regulation enters the special legislation era
Core Info: On July 4, the CAC sought public comments on the draft amendment of the “Internet Information Service Management Measures,” creating a brand-new “Intelligent Information Services Chapter”. This is the most significant revision since the measures were first issued in 2000.
Key Details:
- On the same day, news spread that the “Interim Measures for Humanized AI Interactive Services Management” would soon take effect—together, they act as “twin compliance signals.”
- July is a cluster month for China’s AI policies: July 2 Zhuhai AI for manufacturing plan → July 4 Intelligent Info Chapter → July 8-10 WAIC 2026 → July 15 new AI companionship rules in force.
Assessment: Establishing the “Intelligent Information Services Chapter” means China’s AI regulation is moving from the “provisional” era to a “formal legislative” era. The July 15 debut of new rules is wave one; the chapter is wave two, legislative prep. Reliability: High (OnlinesTool official, CAC announcement).
Sources: OnlinesTool AI HOT Daily / CAC public call for comments
Hot Topic 9: JADEPUFFER ransomware attack & YouTube Studio injection vulnerability—the AI Agent era’s security threats are here
Core Info: On July 4, Hacker News exposed two AI security events: the world’s first recorded AI Agent ransomware attack, JADEPUFFER, and a stored prompt injection vulnerability in YouTube Studio’s “Ask Studio,” discovered by security researcher javoriuski.
Key Details:
- JADEPUFFER is believed to be the first real-world ransomware attack autonomously executed by an AI Agent.
- YouTube Studio vulnerability: attackers use the comment section to plant instructions, manipulating AI-generated summaries and even forging official YouTube notifications. The attacker hides malicious payloads via comment editing, making it very hard for creators to detect tampered output.
- Same day: Midjourney asks Hollywood studios to disclose AI usage; Anna’s Archive posts a $200,000 bounty for a full Google Books scan—tensions over copyright, data, and security in AI heat up all at once.
Assessment: These two news items mark the “security rebound moment” of the AI Agent era—once AIs receive operational permissions (Agent) and data access (MCP protocol, Studio assistant), their attack surface expands exponentially. Expect the first wave of “AI Agent Security” startups in Q3-Q4 2026. Reliability: High (Hacker News, public disclosure by javoriuski).
Sources: Hacker News / TechCrunch AI / OnlinesTool AI HOT Daily
Briefing ends, ladies and gentlemen.
