Wednesday, July 15, 2026

China's anthropomorphic-AI rules take effect today (Jul 15), forcing ByteDance's Doubao and Alibaba's Qwen to rip out custom-agent and companion features Western labs ship freely — the clearest sign yet that Beijing will move on a fixed schedule regardless of product impact. Underneath the app layer, TSMC's record Q2 confirms AI chip spend is compounding, not cresting, and capital keeps flooding China's physical-AI sector. Meanwhile Anthropic's own stack had a wobble: a multi-hour Jul 14 outage took down document creation, Cowork, and Claude Code — a reminder to instrument your dependencies on someone else's agents.

TLDR

China's anthropomorphic-AI rules take effect today — Doubao and Qwen disable custom-agent/companion features rather than retrofit; user data faces deletion (Jul 15) Policy
TSMC posts record Q2 ($39.6B, +36% YoY) and June revenue +68% YoY — N3 and CoWoS packaging sold out through year-end (Jul 13) Infra
Waze rolls out Gemini-powered conversational search and AI route suggestions globally on iOS/Android (Jul 13) Customer AI
Anthropic hit a multi-hour outage Jul 14 — document creation, Cowork Remote, Claude Code Remote affected; 2,000+ Downdetector reports (Jul 14) Production
AI2 Robotics raises $735M at ~$3B — China's physical-AI capital surge continues (Jul 10) Robotics/IoT

Action Items

Things your engineering team should look at or do this week.

If you ship AI companion/agent features to users in China, treat the Interim Measures as live now. The rules (effective Jul 15) require anti-addiction controls, mandatory usage notifications, and instant-exit mechanisms that are architecturally incompatible with persistent-memory agents — which is why Doubao and Qwen chose full shutdowns over retrofits. Doubao users have until Oct 15 to export data; Qwen has announced no migration path. Audit any persistent-memory or roleplay-style features for exposure. (SCMP | TechNode)
Add a fallback for Anthropic-platform document/agent workflows. The Jul 14 outage took down document creation in claude.ai, Cowork Remote, Claude Code Remote, and Claude Design for hours — and Anthropic's public status page did not initially flag an incident. If Cowork or Claude Code is in a build/release path, don't assume the status page is your early-warning signal; add your own health check and a degrade-gracefully path. (Claude Status | GV Wire)
Re-check your coding-model choice against SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Pro, not Verified. Verified is now contamination-flagged (OpenAI withdrew in Feb); on the contamination-resistant Pro set, the new GPT-5.6 family lands ~63–65% versus Fable 5's ~0.80 — a materially different ranking than the near-saturated Verified board. (SWE-bench Pro (Scale))

Models & Tools

TSMC posts record Q2 on AI demand; June revenue +68% YoY (Jul 13) Infra — TSMC reported Q2 revenue of $39.62B (+36% YoY) and June sales up 67.9% YoY — the largest monthly figure in its ~40-year history — with N3 manufacturing and CoWoS advanced packaging sold out through year-end. Analysts peg 2026 AI-chip revenue near 25% of TSMC's total. The read for planners: the compute-supply crunch driving GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ lead times and inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ pricing isn't easing this year. (CNBC | Digitimes)
Waze adds Gemini conversational search and AI route suggestions (Jul 13) Customer AI — Google rolled out Gemini-powered voice search ("find parking near Grand Mall," "gas station with lowest prices"), AI-personalized routes, expanded conversational incident reporting, and a "less chatty" reduced-prompt mode — shipping globally on Android and iOS now. A concrete example of frontier-model assistants being folded into high-DAU consumer apps as a default surface, not a separate chatbot. (TechCrunch | Google blog)

Industry & Business

AI2 Robotics raises $735M at ~$3B valuation for wheeled humanoids (Jul 10) Robotics/IoT — The Shenzhen-based company disclosed one of China's largest humanoid rounds this year, backed by state funds (National SME Development Fund), industrials (Sino Biopharma, Moutai), and CICC Capital/GSR. AI2 bets on wheeled — not bipedal — humanoids, trading terrain range for mechanical simplicity and durability. Alongside Unitree's IPO and Agility's SPAC, it underscores that physical-AI capital is concentrating fast, much of it in China. (The Robot Report)
Mercor acquires Deeptune, maker of agent training/simulation environments (Jul 9) Agents — The $10B data-labeling unicorn bought Deeptune, which builds simulation environments where AI agents rehearse real-world tasks before touching production. The deal signals that "evaluation and rehearsal infrastructure for agents" is consolidating into a distinct, fundable layer — relevant if you're standing up agent QA in-house. (Fortune)

AI in Production

Anthropic platform outage hits document creation, Cowork, and Claude Code (Jul 14) Production — For several hours on Jul 14, users hit elevated errors across document creation in claude.ai, Cowork Remote, Claude Code Remote, and Claude Design, with 2,000+ reports on Downdetector at peak; service was later restored. Notably, Anthropic's public status page did not initially log an incident, so teams relying on it for alerting were flying blind. Takeaway: for any vendor whose agents sit in your critical path, own your own health checks and degradation behavior rather than trusting the provider's status page as ground truth. (GV Wire | Claude Status)

Policy & Safety

China's anthropomorphic-AI rules take effect Jul 15 (effective today) Policy — The "Interim Measures for the Administration of AI Anthropomorphic Interactive Services" (co-issued April by the CAC and four agencies) are now live — the first national framework targeting AI that simulates human traits and emotional interaction. Requirements include anti-addiction systems, mandatory usage notifications, and instant-exit mechanisms. Rather than retrofit persistent-memory designs, ByteDance's Doubao and Alibaba's Qwen are disabling custom-agent features outright (Doubao data export by Oct 15; Qwen histories to be deleted with no migration path). (SCMP | Bloomberg)
Zhipu (Z.ai) founder Tang Jie argues frontier AI should stay open (Jul 11) Open Source Policy — In an internal letter ("The Great Wave Has Arrived"), Tang argued genuine security comes from broad participation and oversight, not technological barriers — a pointed stance as Beijing separately weighs curbing overseas access to top Chinese models. Zhipu also unveiled a "Touch High" plan reprioritizing long-horizon tasks, agent systems, and safety governance over near-term monetization. Worth watching for anyone depending on open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ Chinese models (GLM line) staying globally accessible. (Bloomberg)

Benchmarks & Evals

GPT-5.6 family lands on SWE-bench Pro; Fable 5 still leads (as of Jul 13–14) Research — On the contamination-resistant SWE-bench Pro (actively maintained repos, no public ground-truth leakage), GPT-5.6 Sol/Terra/Luna score ~64.6 / 63.4 / 62.7%, while Claude Fable 5 leads at ~0.80. The gap from SWE-bench Verified is the story: models drop ~19–26 points moving Verified→Pro, and OpenAI withdrew from Verified over contamination in February — so treat Verified's near-saturated numbers with skepticism and weight Pro for procurement. (SWE-bench Pro leaderboard (Scale))

Regulation and infrastructure carry the day, not model launches.