> AI Daily Briefing, Sept 17: AI agents cross into the physical world — Google opens Home to any third-party agent, Anthropic merges Claude chat and Cowork into one workspace, and Agility ships Digit 5, a humanoid built around not hurting people. Meanwhile, Apple is reportedly building its own AI servers on M-series Ultra chips.
## 🔥 Today's Top 3
**1. Google will now let any AI agent run your smart home**
Per The Verge, Google is opening Google Home device control to any third-party AI agent — not just its own Gemini assistant — letting external agents directly operate lights, thermostats and cameras through a standardized protocol. Smart home has long been a walled garden; by handing the entry layer to the agent layer, Google is officially moving AI agents from "answering questions" to "acting in the physical world," and the home is the next battleground.
**2. Anthropic merges Claude chat and Cowork — and takes aim at Google's suite**
Anthropic has unified the Claude chat interface with its Cowork workspace, adding Docs- and Slides-style document and presentation tools. This is no longer "a chatbot with plugins" — Claude is becoming a full productivity suite where conversation, collaboration, docs and decks live in one place. The line between chat entry point and office suite is disappearing, and Anthropic is now competing head-on for the "Office of the AI era."
**3. Agility unveils Digit 5: a humanoid designed around not hurting people**
Agility Robotics launched Digit 5, a new humanoid robot with safety-first engineering — it stops and squats to lower its center of gravity when humans are nearby. While rivals race on payload, speed and price, Agility is selling workplace coexistence to the enterprises that would actually put robots on a factory floor. Commercialization is shifting from "can it work" to "will you let it onto the line."
## 🌐 Global News Digest (8 stories)
- **Anthropic and OpenAI want to embed safety evaluators — but how independent will they be?**: Both labs are moving safety evaluation into model development itself. TechCrunch's question is the hard one: can a lab's in-house evaluators truly be independent? The structural conflict of self-review remains unsolved. Source
- **Apple reportedly building an AI server packed with M-series Ultra chips**: Ars Technica reports Apple is building AI servers around its M-series Ultra silicon, aiming to cut Nvidia dependence for inference infrastructure. Apple re-entering server hardware means its in-house chip stack is finally pointed inward. Source
- **Meta readies camera-less smart glasses after 'perv glasses' backlash**: Following accusations that its Ray-Ban glasses enable covert filming, Meta is preparing a version without a camera. The privacy controversy is big enough that Meta will drop a core feature — wearables' boundaries are being redrawn by public opinion. Source
- **AI marketing startup Profound raises $180M Series D at $1.8B valuation**: Per 36Kr, answer-engine-optimization startup Profound closed a $180M Series D at a $1.8B valuation — just months after its unicorn round, as "getting cited by AI" keeps heating up. Source
- **27-year-old Tsinghua professor raises big funding in two months, betting alongside Jeff Dean**: Per 36Kr, a 27-year-old Tsinghua assistant professor raised hundreds of millions of yuan within two months of founding a startup, in a direction aligned with Google legend Jeff Dean's bet. Star academics founding frontier labs is the new capital narrative in China's AI scene. Source
- **ZDTaichu5.0-9B: an open Chinese multimodal model joins the global first tier**: Per QbitAI, a USTC-HFUT team open-sourced ZDTaichu5.0-9B, claiming full general capability with standout spatial embodied intelligence and a place in the global multimodal top tier. Small models punching up on multimodal benchmarks strengthens the case for edge deployment. Source
- **Bilibili launches an AI arena: 100+ models compete in public**: Per QbitAI, Bilibili's AI arena went live with over a hundred global models competing head-to-head, GPT-6 on top. Model arenas are moving from academic benchmarks to consumer platforms where anyone can compare models directly. Source
- **The AI data center e-waste problem is huge — and getting bigger**: The Verge reports that AI data centers are generating enormous and fast-growing amounts of e-waste, as rapid GPU iteration retires hardware early. AI's environmental bill is expanding from power draw to physical disposal. Source
## 📊 Tech & Trend Watch
- **Agents move from chatting to taking over**: Google opening Home to third-party agents, Snap's Specs agent tooling, and Anthropic merging Cowork all point the same way — agents are becoming OS-level entry points, and platforms are racing to own the default "do it for you" position.
- **The hardware shadow war over AI infrastructure escalates**: Apple's Ultra-chip AI servers and reported SK Hynix–Intel talks on US memory production signal accelerating vertical restructuring of the compute supply chain — "owning your full chip stack" is the new contest.
- **Safety and governance enter deeper water**: Questions over the independence of lab-run evaluators and rising attention to data center e-waste mean the costs of AI expansion — governance gaps and environmental externalities — are being put on the agenda one by one.
---
Disclaimer: For reference only; exercise caution for investment or decisions.