The US flagship started in Beijing

Welcome back technologists🫡
(ayo tyler denk founder of beehiiv here’s ur intelligence report)

Writer priced a Fortune 500 flagship at $2 in and $8 out per million tokens. The starting weights came from Z.ai in Beijing. Kobe did not care. I did.
(about 4 minutes)

Here's the signal today: 🛸

👨‍💻🐕️

🧬 US enterprise flagship is a post-trained Chinese open model
🛠️ One command wires every coding agent, Exa free through August 31
🛠️ GLM-5.2 free on eve through August 27; pin DeepSeek's dated ID
🛠️ 53% already had an agent incident; 8% isolate and enforce
✂️ Four Copilot features die August 18
✂️ Meta shipped a free Creator Studio app for iOS
🌴 California's billionaire tax has a January 1 residency hole

🧬 Writer's Palmyra X6 is GLM-5.2 with a US post-train

A Fortune 500 flagship just admitted its weights started in Beijing. Writer released Palmyra X6 on August 13. It says the agent product now runs 52% cheaper, 48% faster, and 10% better. The model is not trained from scratch. It is a post-trained GLM-5.2, the open-weight mixture-of-experts model from Beijing-based Z.ai. Writer prices X6 at $2 / $8 per million tokens against Opus 4.8 at $15 / $75. Accenture, Uber, and Vanguard are named customers.

The trap: a US vendor and US infrastructure mean a US model. They do not. Matan-Paul Shetrit, Writer's director of product management, told VentureBeat the model "is in no way, shape, or form connected to any of its original developers" and "is fully run on our U.S. infrastructure." SaferAI found raw GLM-5.2 refused none of the offensive cyber or biology tasks it was given on Z.ai's public API. Artificial Analysis scored GLM-5.2 at 51 on its Intelligence Index.

Money play: post-train an open-weight model before you renew a frontier seat. Price the fine-tune against the API bill. If you ship the result, keep the provenance file. Customers will ask.

🛠️ Agent Watch

One command, every agent. Vercel's August 13 changelog wires Claude Code, Codex, OpenCode, and more to AI Gateway with one CLI command. @ai-sdk/harness-acp lets any Agent Client Protocol setup plug into the same agent interface. Exa web search works as a tool with any Gateway model, free through August 31, no Exa key. GLM-5.2 is free for eve agents through August 27 via Blackbox. DeepSeek V4 Pro now serves updated weights by default; the dated model ID pins the version.

The trap: "free this month" is not a stack. It is a window. If you do not pin the dated ID, a silent weight update diffs your evals.

Run the CLI setup and pin the dated DeepSeek ID before August 27. Use the free Exa window to test search-augmented agents, then decide if the bill is worth keeping.

Runtime permissions are not containment. Visa's Rajat Taneja walked VB Transform 2026 through aiming Anthropic's Mythos at Visa's own payment network. The model stitched minor weaknesses into working exploit chains. Visa open-sourced the setup that governed the hunt. VentureBeat Pulse, 440 enterprise security respondents: 53% already had an agent incident or near-miss. 65% enforce permissions at runtime. 18% isolate their highest-risk agents. 8% do both.

If your riskiest agent can touch payments, email, or production data, sandbox it this week. Identity without isolation is how the other 92% live.

✂️ The Build

Four Copilot features die August 18. Microsoft is merging consumer and business Copilot into one app. Official support: Group Chat threads, messages, and images are not carried forward after August 18, 2026. Podcasts stop. Deep Research retires for consumers the same day. GeekWire, August 13: Mico, the voice blob, goes with them. Researcher, the Deep Research stand-in, is a Microsoft 365 Premium feature, not Personal or Family. Work and personal accounts stay separate. Data does not flow between them.

The trap: the brand stayed. The room you were renting did not. Export podcasts and Group Chat images before Tuesday. If Deep Research is in the workflow, you need Premium or a different tool, not hope.

Meta's free creator desk. Facebook rolled out a stand-alone Creator Studio app on iOS on August 12 for creators in the US and Canada. The assistant answers "When should I post?" and "What are people saying in my comments?" from your style, performance, and goals. It can draft comment replies in your tone. You still approve before they post.

If you publish on Facebook, install it and ask the comment-sentiment question before the next post. Free analytics that already has your numbers beats pasting insights into a consumer chatbot.

🌴 Continental

The residency snapshot is the tax. The California Billionaire Tax Act, if voters approve it in November, would impose a one-time 5% tax on the net worth of anyone who was a California resident on January 1, 2026, with wealth measured December 31, 2026. The Tax Foundation's Jared Walczak: because the measure cannot be enacted until early November, its residency snapshot predates the tax by ten months. Billionaires who leave in 2026 have a live challenge. Zuckerberg bought a Florida mansion in February.

If you have California exposure, the file that matters is January 1, not the day you file. Document move date, voter registration, and lease as of that snapshot.

🎓 Level Up

Visa's number is the one to act on: 65% enforce at runtime, 8% pair that with isolation. Four lines, one agent, twenty minutes:

AGENT CONTAINMENT AUDIT (run per agent, not per system)

1. IDENTITY: Does this agent have its own credentials,
   or is it borrowing a human's? (Borrowed = fail.)
2. ISOLATION: Can this agent's failure touch production
   data, payments, or email? If yes, sandbox it.
3. BLAST RADIUS: List every tool the agent can call.
   Remove any tool no task in the last 30 days used.
4. KILL TEST: Revoke its credentials at 2pm on a Tuesday.
   Did anything outside the sandbox notice? Fix what did.

Score: 4/4 = contained. Anything less = the 92% that
do not isolate their highest-risk agents.

Run it against your riskiest agent before Friday.

Reply and tell me: are you running any agent on Chinese open weights in production, and did your customers ask? I read every reply.

Chase (and Kobe, who contains exactly zero agents)
Nomad Signal | nomadsignal.ai. Forward this to one builder who needs it.

See you tomorrow,

Chase & Kobe 👨‍💻🐕

One tap. Feedback goes straight to the desk.