Beijing will not pace the frontier, Google's quantum error breakthrough, an AI that maps immune cells, routers as AI plumbing, and agent tracing on Claude
Hello, dear TEA-mates! Here is what you need to know today.
1. 🇨🇳 SemiAnalysis: Beijing Will Not Pace the Frontier
After Dario Amodei's September 12 essay urged frontier labs to slow capability gains, SemiAnalysis examined what China actually does on AI safety. Beijing's new AI Safety Governance Framework 3.0 (September 14) warns about recursive self-improvement and models that deceive evaluators, yet it names "promoting AI innovation and development as the first priority," and the State Council targets 70% penetration of AI agents and intelligent terminals by 2027 and 90% by 2030. China's rules govern content and applications, with no duties triggered by training compute or capability. Of 857 model releases from nine leading Chinese developers, only 31 (3.6%) ever came with a published safety result, and just 9 (1.1%) at launch. Zhipu is the only lab with results every year since 2022, and state-backed evaluators have no power to demand access to models. (Read More)
🫖 TEA For Thought: "I hope this really wakes up those who talk about slowing down AI. AI is now as important as nuclear, if not more important. The dictatorship isn't going to stop, regardless of whether the West slows down or not. In fact, they couldn't beg more for the democracies to slow down so they can speed up to the maximum and dominate the world."
2. ⚛️ Google Cuts Magic-State Errors Roughly 40-Fold on Willow
Google's Quantum AI team reported in Science that a technique called "magic state cultivation" reduced magic-state errors roughly 40-fold compared with uncultivated states on its Willow processor. Magic states let quantum computers run the operations needed for universal quantum computing, but producing them cleanly is expensive: some estimates say it could consume up to 95% of a machine's capacity. Google's method uses coordinated measurements to catch errors and discard faulty attempts, reaching 99.99% logical fidelity for T-states with fewer resources. Researchers also moved cultivated states between two error-correction methods, a step toward fitting the technique into Google's existing architecture. The experiment confirms an approach first proposed in theory in 2024, though Google says more work is needed to scale it. (Read More)
🫖 TEA For Thought: "Google and IBM might be leading more than you can ever imagine."
3. 🧬 New AI Tool Shows What Draws Immune Cells to Tumors
Researchers at The University of Texas MD Anderson Cancer Center and partner institutions built SPACER (Spatial Analysis of Cellular Engagement and Recruitment), an AI model that explains why immune cells gather in some tissue areas and avoid others. Published in Nature Biomedical Engineering and co-led by Tao Wang and Linghua Wang, it reads spatial transcriptomics data and studies groups of neighboring cells rather than isolated ones, so its predictions can be explained. Across 37 datasets from several cancer types, it was 4 to 21 times better than existing approaches at finding meaningful recruitment signals. It found that mucins, sugar-coated proteins many cancers overproduce, may form a barrier that keeps T cells out, and high mucin levels tracked with poorer responses to checkpoint inhibitors in melanoma and gastric cancer. (Read More)
🫖 TEA For Thought: "Isn't this beautiful!"
4. 🔀 Vals AI: Routers Are the New AI Primitive
Evaluation firm Vals AI argues that model routers, which pick a model for each request based on quality, cost, and latency, are becoming a core layer of the AI stack. On the Vals Index, five different models hold the top spots across eight benchmarks, and none ranks in the top three on all of them: Claude Sonnet 5.5 and Claude Opus 5.5 rank second and third overall but 35th and 33rd on Harvey's Legal Agent Benchmark, which Meta's Muse Spark 1.2 tops. Ten developers released 16 language models in the past month alone. A perfect "oracle" router would beat the best single model by 14 points while costing about 40% less. Vals says most academic router benchmarks test isolated public questions, so it is building long-horizon agent evaluations and will publish a public router leaderboard soon. (Read More)
🫖 TEA For Thought: "Routers will be the most basic infrastructure for all companies."
5. 🔭 Claude Platform Agents Now Send OpenTelemetry Traces
Anthropic now lets developers send OpenTelemetry (OTel) traces from Claude Managed Agents, the Messages API, and Claude SDKs to any OTel-compatible tool, now in public beta. After adding a destination in the Claude Console, every agent turn, model call, tool use, and MCP server call arrives as a standard trace with latency, tokens, and IDs. Managed Agents and the Messages API need that destination set, while Claude SDKs export traces automatically. Traces follow the OpenTelemetry GenAI conventions, so Claude data sits beside other providers in Grafana Cloud, Honeycomb, Sentry, Splunk, Datadog, Arize AI, Braintrust, Langfuse, LangSmith, New Relic, PostHog, and Databricks. Anthropic's pitch: review an agent's traces, then widen its access as it proves reliable. (Read More)
🫖 TEA For Thought: "Observability plus verification might be becoming a standard for all agentic systems."
🛠️ Skill of the Day
The Monday Reset: Turns last week's messy notes into an honest review and a short, realistic plan for the week ahead.
Act as a calm, practical chief of staff. Help me review last week and plan this one.
Here is what happened last week: [PASTE NOTES, CALENDAR, TO-DO LIST, OR A QUICK BRAIN DUMP].
Here are my goals for this month: [LIST 1 TO 3 GOALS].
Hours I realistically have this week for focused work: [NUMBER].
Then do the following:
1. Wins: list what actually moved forward, in one line each.
2. Stuck: list what stalled and the most likely reason (unclear next step, waiting on someone, too big, or not important).
3. Drop: name anything I should stop doing or hand off, and why.
4. This week: pick no more than 3 priorities that fit my available hours, each with a concrete first step I can do in under 30 minutes.
Rules: be honest, not cheerful. Use only what I gave you. If my plan does not fit my hours, say so and tell me what to cut.Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own notes, goals, and hours.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: