> AI Daily Briefing, Sept 16: AI safety moves from war of words to the negotiating table — OpenAI, Anthropic and Google have reportedly been in secret talks for weeks; Meta puts AI agents inside WhatsApp Business; AEO startup Profound becomes a unicorn with a $180M raise while DeepSeek gets a new CFO.
## 🔥 Today's Top 3
**1. OpenAI, Anthropic and Google have been in talks on AI safety for weeks**
After a string of AI leaders publicly called to slow development — and were rebuffed by the White House and Nvidia — TechCrunch reports that OpenAI, Anthropic and Google have quietly negotiated on AI safety for weeks. Rather than fighting it out in public, the three labs appear to be trying to draw their own safety red lines before regulators do. The "pact or cartel" debate is heading toward a concrete answer.
**2. Meta now lets AI agents handle the boring parts of WhatsApp Business setup**
Meta rolled out AI agents that automatically handle the most tedious parts of WhatsApp Business setup — catalog building, greeting messages, auto-reply rules. It's Meta's latest step from chatbot assistant toward "do-it-for-you" agent, and a sign that every major platform is racing to embed agents as default infrastructure for small-business automation.
**3. DeepSeek gets a new CFO — an investor who backed Zhipu and MiniMax**
Per QbitAI, DeepSeek has brought in a new CFO with a track record of investing in Zhipu and MiniMax. With Anthropic reportedly eyeing a $2-trillion IPO and fundraising across the industry accelerating, the hire reads as a signal that the famously cost-efficient "efficiency camp" is preparing for a bigger capital game.
## 🌐 Global News Digest (8 stories)
- **AEO startup Profound hits unicorn valuation with $180M Series D, 7 months after last round**: The post-SEO "answer engine optimization" market just produced a new high-water mark — Profound raised $180M seven months after its previous round, as enterprises rush to optimize for being cited by AI engines. Source
- **Nvidia open-sources its IMO gold-medal recipe — with a 1.5TB-VRAM catch**: Leiphone's teardown of Nvidia's open-sourced IMO (math olympiad) technique argues it's not just "brute force": massive verification compute and a 1.5TB VRAM cluster reset the entry bar for frontier open-source AI. Source
- **InferenceX, Tsinghua and SJTU open-source APXInf, an on-device embodied inference engine**: A team led by InferenceX with Tsinghua and Shanghai Jiao Tong University released APXInf, an open-source inference engine for embodied AI on edge devices, claiming SOTA on Pi 0.5-class tasks — key infrastructure for running embodied models without the cloud. Source
- **Zhiseng releases a physics-grounded video model, joining the global first tier**: Zhiseng (ZX) unveiled a video generation model oriented around physical laws, which it says places it in the global top tier of AI video. Physical consistency, not prettier pixels, is becoming the new dividing line in video generation. Source
- **Exclusive: Paying for frontier AI models buys a 4-month head start at 5x the cost**: Ars Technica's exclusive comparison finds Silicon Valley's frontier closed models lead open Chinese models by roughly four months — at five times the cost, a gap that makes "subscribe to the frontier" much harder to justify on economics alone. Source
- **Agility's new humanoid robot will stop and squat to avoid harming human coworkers**: Agility Robotics' next-gen humanoid adds active avoidance — halting and lowering its center of gravity when humans are near. Humanoids are moving from "can work" to "can share a floor," making workplace safety a core engineering focus. Source
- **AI and data centers are incredibly unpopular in every poll**: The Verge rounds up polling showing AI and data centers score poorly with the public across the board. With midterms approaching, grassroots resistance to AI infrastructure is becoming a policy variable. Source
- **Meta's new One subscriptions put a price on social media and AI**: Meta launched "One," a subscription tier bundling social perks with AI capabilities — formally making "pay for AI" a second growth curve alongside advertising. Source
## 📊 Tech & Trend Watch
- **AI safety moves from the spotlight to the back room**: After Huang publicly refused a slowdown and the White House called the pause chorus "overreacting," the three biggest labs reportedly negotiated in private — the appearance of a split masks a rapid coordination of positions, and the window for self-regulation ahead of legislation is closing.
- **Capital and talent are flowing to monetization, not models**: Profound's unicorn round, DeepSeek hiring a CFO who backed Zhipu and MiniMax, Meta embedding agents into merchant tooling — the race is shifting from "who has the best model" to "who can charge for it," and investors are betting on the revenue infrastructure.
- **Open source is closing the gap quarter by quarter**: Nvidia's open-sourced IMO recipe, an open on-device embodied engine from InferenceX/Tsinghua/SJTU, and Ars' four-month lead estimate all point the same way — the price premium for frontier closed models is getting harder to defend.
---
Disclaimer: For reference only; exercise caution for investment or decisions.