Waymo's Chinese imports, Stanford's widening youth employment gap, NVIDIA's model router, GitHub on orchestration, and AI agents that simulate A/B tests
Hello, dear TEA-mates! Here is what you need to know today.
1. 🚕 Waymo Imported 3,200 Chinese Robotaxis Despite 102.5% Tariffs
Alphabet's Waymo has cumulatively imported more than 3,200 Zeekr-based Ojai robotaxis into the United States despite 102.5% tariffs, according to Forbes reporting on customs records. More than 2,600 of those shipments were recorded in 2026, and over 500 Ojai vehicles were photographed together at Waymo's integration facility in Mesa, Arizona, in August. The Ojai is built by Zeekr in Ningbo, China, on Geely's SEA-M architecture, and carries an 800V system, a 93 kWh battery, and a 200 kW rear motor. Customs records do not name Waymo as the consignee. Chinese factories supply the body, battery, and drivetrain, while Waymo installs its sixth-generation Driver (four LiDAR sensors, six radars, 13 cameras) in the US. Import declarations value the CM1e chassis at roughly 38,000 to 38,500 USD, putting the landed chassis near 78,000 USD after tariffs, or about 103,000 USD once Waymo's disclosed 25,000 USD hardware is added. That is roughly 48.5% below the estimated 200,000 USD cost of its Jaguar I-Pace generation. (Read More)
🫖 TEA For Thought: "Would you dare to ride an autonomous vehicle manufactured in China? Americans have no idea what the CCP is capable of."
2. 📉 Stanford Finds The Young-Worker AI Gap Has Widened To 19%
Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen released a revised version of "Canaries in the Coal Mine?", using ADP payroll data to track employment since ChatGPT's release. They report six facts. There is no widespread economy-wide displacement, but employment among workers aged 22 to 25 in highly AI-exposed occupations now sits about 19% below where it would be had it kept pace with less-exposed peers, up from 15% at the July 2025 vintage. In levels, that age group fell about 11% in the two most exposed quintiles between November 2022 and June 2026 while growing about 10% in the three least exposed. Experienced workers show no comparable gap. The adjustment runs through reduced hiring rather than layoffs, and it concentrates in occupations where AI automates rather than complements. The new revision separates codified knowledge, which is declining for the young, from tacit knowledge, which is holding up. The authors call these descriptive patterns, not causal estimates. (Read More)
🫖 TEA For Thought: "How do young people adjust and adapt to this rapid AI era, emotionally, mentally, and capability-wise? When higher education is basically useless now, with expensive tuition and an outdated curriculum, and the employment rate is low especially for knowledge workers, where is the way out?"
3. 🔀 NVIDIA Open-Sourced A Router That Swaps Models Mid-Task
NVIDIA published NeMo Switchyard on August 11, 2026, a fully open-source library that routes agent workloads across specialized and frontier models at runtime, judging each request on capability, cost, latency, and system signals without requiring teams to rebuild their applications. The library routes across a mixed field including DeepSeek V4, Kimi K2.6 and K2.7, Qwen3.5 397B A17B, Claude Opus 4.8 and Opus 5, and NVIDIA's own Nemotron models. LangChain reported that routing between Nemotron 3.5 Lightning and Claude Opus 4.8 with the escalation router produced a 74% cost reduction against a frontier-only baseline across five runs, while sending just 7% of calls to the frontier model, at a measured 6-point accuracy tradeoff. Cognition reported 50.6% accuracy at a 3.11 USD mean cost, landing within 2.8 percentage points of Opus 5 at roughly 28% lower mean cost. NVIDIA says the pairing holds frontier-level task completion at about a third of the benchmark cost of Opus 4.8 alone. (Read More)
🫖 TEA For Thought: "Switchyard is something enterprises would need when it comes to budget control."
4. 🎛️ GitHub Says The Developer's Job Is Now Designing The System
GitHub argues that the one-prompt demo is the easy part and that the real work is wiring a system that generates code reliably and safely, which shifts the developer from writing code to designing how code gets proposed, validated, reviewed, and shipped. The recommended pattern starts from ordinary repository events, a label added to an issue or an overnight scheduled workflow, which trigger a GitHub Actions workflow that invokes an agent on a scoped task. The agent's output lands in a pull request, where deterministic checks take over: linting, tests, security scanning, and build verification, with CODEOWNERS, required reviews, and branch protections governing what can merge. The argument is that agents handle ambiguous, context-heavy work inside a rule-based boundary, and that the deterministic side is what makes teams trust the system. GitHub advises starting small with one bounded workflow such as issue triage or docs-and-tests sync. (Read More)
🫖 TEA For Thought: "Very well put."
5. 🧪 Can AI Agents Predict Your A/B Test Before You Run It?
Stefan Hut and Lorenzo Masoero propose the Simulated Randomized Controlled Trial (S-RCT), a framework for testing whether AI agents conditioned on behavioral profiles can predict experiment outcomes before real traffic is spent. They derive a two-layer error decomposition that separates agent approximation error from subsampling error so each can be improved on its own, and the framework is agent-agnostic, accepting any behavioral model from a fine-tuned specialist to a general-purpose foundation model. Validated against 67 historical marketing A/B tests, a baseline S-RCT built on an off-the-shelf foundation model captured directional signal with a sign overlap of 0.70 but systematically overshot effect magnitudes. A two-phase pre-period calibration protocol cut squared prediction error, after removing irreducible measurement noise, by roughly 77 times. A within-subject design exposing each agent to both arms reduced standard errors by about 2.4 times. (Read More)
🫖 TEA For Thought: "The paper is a workshop-accepted preprint, so its strong calibration gains should be treated as promising but not yet general proof across all consumer-product experiments."
🛠️ Skill of the Day
The Tacit Skill Finder: separates the parts of your work a model can already copy from the parts it cannot, and tells you what to build next.
You are a career strategist who studies how work is actually
done, not a motivational coach. Be blunt and specific.
My role: [YOUR JOB TITLE]
What I actually do in a normal week: [LIST 6 TO 10 TASKS,
PLAIN LANGUAGE, INCLUDING THE BORING ONES]
Years of experience: [NUMBER]
Tools I already use: [LIST THEM]
Sort every task I listed into three buckets:
1. Codified. The knowledge is written down somewhere, so a
model can reproduce it from text. Say how exposed it is.
2. Tacit. It depends on judgment, relationships, context, or
knowing which rule to break. Explain what makes it hard
to copy.
3. Mixed. Say which half is which.
Then give me:
- The two tacit strengths I should deliberately deepen, and
what "deeper" concretely looks like in six months.
- The one codified task I should hand to a tool this month,
and what I should do with the time it frees.
- One blind spot: something I called tacit that is probably
codified, and why.
Do not reassure me. If most of my week is codified, say so
directly and tell me what the fastest route out looks like.Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: