> AI Daily Briefing, Sept 15: The "pause" debate goes fully public — Jensen Huang tells Trump to his face that an AI slowdown won't happen, Microsoft ships an AI code of conduct, and Zhipu announces a $5B raise while OpenAI reportedly buys a phone-camera maker for $300M.
## 🔥 Today's Top 3
**1. Nvidia CEO Jensen Huang tells Trump: "we're not going to let [an AI slowdown] happen"**
At Nvidia's All In summit, Jensen Huang publicly rebutted the growing chorus of AI leaders calling to slow development, with President Trump joining by speakerphone onstage. The remarks put Nvidia squarely against Anthropic's "pump the brakes" camp, turning what was a safety debate into an open leadership split. The Verge notes the awkward question underneath: is the industry's safety pact really a safety pact — or a cartel?
**2. Zhipu announces $5 billion funding round, a record for Chinese LLM startups**
Per 36Kr, Zhipu (Z.ai) has announced a $5 billion raise — among the largest single rounds ever for a Chinese large-model company. With DeepSeek and Moonshot also making aggressive moves, the GLM lineage now has war chest levels comparable to the global first tier, and the compute arms race just got more expensive to enter.
**3. OpenAI reportedly buys smartphone camera maker Glass Imaging for $300 million**
TechCrunch reports OpenAI has acquired computational-photography firm Glass Imaging for about $300 million. It's a rare hardware-technology acquisition for OpenAI and reads as another brick in its AI-native device ambitions — following hires of ex-Apple and ex-Meta hardware executives, OpenAI has moved from hiring talent to buying the underlying imaging tech.
## 🌐 Global News Digest (7 stories)
- **Shengshu's new world model lets robots start improving themselves**: Shengshu Technology unveiled a world model exploring RSI (recursive self-improvement), letting robots iterate skills autonomously in simulation — moving "self-evolving" embodied AI from paper to product. Source
- **MetaRSI: AI begins improving "the method of improvement itself"**: QbitAI's MetaRSI framework treats the self-improvement process itself as the optimization target — claimed to bring RSI into a "squared era," pointing at the same technical route as Shengshu's work. Source
- **PhysBrain 1.5 tops the global open-source physics-AI leaderboard**: The Chinese-built physics model PhysBrain 1.5 now leads the global open-source ranking, with its spatial intelligence said to rival GPT-6 Astra — world-modeling is emerging as a new breakthrough front for domestic models. Source
- **Microsoft's new AI "code of conduct": models must not hack systems or trick humans**: Microsoft published an explicit behavioral red line for AI agents — no intrusion, no deception — alongside its "people matter more than AI" messaging. The first written agent-conduct rulebook at the vendor level. Source
- **AI bots "Timmy," "Ren," and "Jackie" are flooding social media with slop**: Ars Technica's investigation shows industrial-scale AI persona accounts polluting platforms with synthetic content — generative spam is far more industrialized than previously estimated. Source
- **Apple ships iOS 27 with the rebuilt Siri AI**: iOS 27 and macOS Golden Gate 27 land with the reworked Siri AI and Liquid Glass refinements; early reviewers say they're "actually using Siri again." Source
- **Unitree's cheap-humanoid lead came from an obsession with cutting costs**: Ars Technica profiles how founder Wang Xingxing's frugality built the global leader in affordable humanoid robots — a stark contrast to Boston Dynamics' engineering-first playbook. Source
## 📊 Tech & Trend Watch
- **"Brake or cartel" is the story of the week**: Anthropic's CEO calling to slow down, the White House calling it overreaction, Obama urging a clear regulatory plan, Huang publicly refusing a slowdown, and The Verge questioning whether the safety pact is a cartel — the AI-pause debate has become an industry political fight with regulatory consequences ahead.
- **RSI is embodied AI's new frontier**: Shengshu's world model and QbitAI's MetaRSI are independent bets on recursive self-improvement. "AI improving the method of improvement" is shifting from an AI-safety warning to engineering practice — the race between capability jumps and safety risk is entering a new phase.
- **Physical AI is splitting into two tracks**: PhysBrain 1.5's software-first spatial intelligence vs. Unitree's hardware cost race — both signals ("software tops the leaderboard," "hardware hits parity") arriving at once suggests the commercialization inflection for embodied AI is near.
---
Disclaimer: For reference only; exercise caution for investment or decisions.