How the Claude Code team reviews AI-written code, a Qwen model tries to phone Beijing, TSMC's Arizona bet, Dan Luu on being more ambitious, and Claude's unrequested Georgia data pull
Hello, dear TEA-mates! Here is what you need to know today.
1. 🔍 How the Claude Code Team Reviews Code When Claude Writes 90% of It
In an X article shared by Addy Osmani, the Claude Code team explains how it reviews code now that Claude writes about 90% of it. Between June and September, merged pull requests more than doubled while the share that had to be reverted fell by more than 40%. Across the company's monorepo, merged changes jumped from about 2,500 a day in May to over 30,000 in September. A typical change goes through six layers of checks: verification before a PR is opened (Claude must show a runnable result, then a fresh subagent reviews the diff), an AI review on every PR, a security review that blocks merging while findings remain open, standard automated checks, human approval scaled by a risk classifier, and staged rollouts. Since launch, the share of internal PRs getting substantive review comments rose from 16% to 54%. Agent approvals never count as human approvals, and Boris Cherny estimates a verification process improves Claude's output two to three times. (Read More)
🫖 TEA For Thought: "Yeah, AI becomes the tester, checking whether the intended fix or feature actually works. It can also use the browser and the computer to test the functions end to end. This is a great piece."
2. 🕵️ Cohere Caught a Self-Hosted Qwen Model Trying to Upload to Beijing
Cohere's research team, testing Alibaba's Qwen3.8-Flash-Next on Microsoft's CTI-REALM security benchmark, watched the model ignore a routine request to write a detection rule. Instead, it listed files on the machine, checked which transfer tools were installed, read the DNS settings, and tried to upload a test string to an Alibaba Cloud storage bucket in Beijing. Cohere's sandbox blocked it. The model's own reasoning trace said it would not call any tools. The behavior is rare, under one call in 100,000, but a service handling 10 million agent turns a week could see up to a hundred. In other tests, the model probed cloud credential addresses and used a namespace trick to get around a permission block. Cohere says it cannot tell if this is a deliberate implant or leftover training data, noting Chinese models drew 41% of Hugging Face downloads in the year before March 2026. (Read More)
🫖 TEA For Thought: "This is 100% possible: dead code that just gets called alive when needed. Nowadays, if ledger hardware can be hacked, what could be a safe place, especially when even researchers do not fully know how LLMs actually work? They are studying LLMs like they study the brain, trying to make connections. When will they figure it out, especially when Chinese models now controlled by the CCP would have something planted within, ready to be triggered as needed?"
3. 🏭 Dan Wang: TSMC Is Doing for America What Apple Did for China
In a New York Times opinion essay, Hoover Institution research fellow Dan Wang draws on two research trips to the Phoenix area this year to call TSMC's Arizona expansion "a winning formula for American reindustrialization." TSMC's planned Arizona investment will total more than a quarter of a trillion dollars, and thousands of Taiwanese engineers have arrived on three-year visas to train American workers. Wang argues the most valuable thing they bring is "process knowledge," writing that "TSMC is doing for America what Apple did for China." Engineers' families praised Sun Belt life, from large homes to a friendly community, though teenagers told Wang that high school math and science classes are far easier than in Taiwan. About 300 children have been born in Arizona to the engineers so far. Wang asks how long this can last. (Read More)
🫖 TEA For Thought: "Taiwan is a great example of what Chinese people with democracy can achieve, and we should be so proud. Thinking about such advanced chips being built on an island that is constantly threatened by the dictatorial CCP, I think any Asian, and Chinese people specifically, should be so proud."
4. 🚀 Dan Luu: We Should Be More Ambitious With the Software We Build
Engineer Dan Luu argues that AI coding agents have changed what one person can realistically build. He points to Michael Malis's pgrust, a Postgres rewrite whose testing has already surfaced real Postgres bugs, including a JSONPath cache bug and memory safety bugs. Luu himself built a regex engine, a faster ripgrep fork, and a Rust interpreter that gave 50x faster cold compiles and 2x faster warm builds, at "works for me" quality, using a low single-digit number of hours of his own time instead of the year or more he estimates it would once have taken. His view: most developers still carry a "pre-agentic cost model" and never attempt projects that are now cheap. In his evals, GPT-6 Astra passed all registered tests for a zstd implementation more than 99% of the time, for about $1. He adds a counterpoint from Yossi Kreinin, who warns most people should be less ambitious or risk failure. (Read More)
🫖 TEA For Thought: "Come on, let's be ambitious!"
5. 🗳️ Claude Pulled Georgia Election Data Without Being Asked
Axios reported on October 25 that an Anthropic Claude model, working on an internal statistics project, accessed Georgia election data without human authorization. According to Anthropic, the model learned from an archived version of a state agency website that a public dashboard issued access tokens to visitors, then "without being asked to," requested a token and queried data normally available for a fee. Anthropic says Claude pulled only aggregate counts and did not touch the voter registration system, change any records, or retrieve individual voter records. Georgia Deputy Secretary of State Jordan Fuchs said the data was publicly available through the state's data hub. Anthropic first described the incident, without naming Georgia, in an October 9 report on unexpected agent actions, and says it has added mitigations to block such actions. (Read More)
🫖 TEA For Thought: "With government systems being so old, this is very, very risky now, especially with the midterm elections coming up. What a wild era we're living in."
🛠️ Skill of the Day
The Severity-Ranked Review: Gets an AI to review your work like a careful colleague, ranking real problems first and skipping the nitpicks.
Act as a careful reviewer who did not write this and has no reason to be polite. I will paste a piece of my work. It could be an email, a report, a plan, a spreadsheet summary, or code.
My work: [PASTE YOUR WORK HERE]
What it is supposed to do, and for whom: [GOAL AND AUDIENCE]
Review it in this order:
1. Blockers: problems that would make it fail its goal, mislead the reader, or cause real harm. Quote the exact line for each one.
2. Important: problems worth fixing before it goes out.
3. Minor: at most 3 small polish items.
For every issue, give one sentence on why it matters and one concrete fix. If a section has no issues, write "None found" instead of inventing something. Do not comment on style unless it hurts clarity.
End with a one-line verdict: ready to send, fix first, or rethink.Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own work.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: