Apple's China-only model, App Store exit fees, the AI hiring divide, software factories, and an autonomous hack on Taiwan
Hello, dear TEA-mates! Here is what you need to know today.
1. 🍎 Apple Trained a China-Only AI With Alibaba
Apple has trained a large language model built specifically for the Chinese market with support from Alibaba, Reuters reported Friday, a shift in how the company approaches AI in a country where OpenAI's ChatGPT and Anthropic's Claude are unavailable. China's Cyberspace Administration registered Apple's generative AI service in July, clearing the main regulatory hurdle for an Apple Intelligence launch in the coming months. Under the approved arrangement, Alibaba's Qwen model is expected to be integrated across compatible iPhones, iPads, Macs and Vision Pro devices in China, and Apple has already let eligible Mac users in the mainland connect Qwen to Siri and Writing Tools. The exact role of Apple's own China model alongside Qwen remains unclear. The stakes are commercial: IDC put China's Q2 2026 smartphone shipments down 4.3% year on year at about 66 million units, with Apple one of only two major makers to grow, up 24.4% to an 18.1% share behind Huawei's 22.6%. (Read More)
🫖 TEA For Thought: "Apple is literally dancing with the evil CCP."
2. 💳 Apple Puts a Price on Leaving the App Store
Apple has filed its proposal for what it should be allowed to charge on purchases made outside the App Store, submitting the plan hours after the Supreme Court denied its request to pause the lower-court proceedings in the Epic Games case. The proffer asks for 15% on standard apps (which pay 30% through in-app purchase), 10% for the Video, News and Mini Apps Partner Programs and for subscription renewals, and 5% for Small Business Program apps. Apple argues that at these rates a large share of US developers can link out profitably, creating competitive pressure on its own payment system while still compensating it for its tools and technologies. The company benchmarked its numbers against Google Play's linked-out rates of 20% standard, 15% program and 10% subscription, terms Epic itself agreed to. Judge Yvonne Gonzalez Rogers is weighing whether Apple's earlier 27% charge was contempt. Epic responds next, and Apple's Supreme Court brief is due September 14. (Read More)
🫖 TEA For Thought: "When a company grows as big as Apple, charging a fee becomes a societal question about how we define monopoly. After all, what is needed is not equality of outcome but equality of opportunity."
3. 💸 Your Interview Now Screens for a $100 Subscription
Engineering interviews increasingly test fluency with paid agentic coding tools, and LeadDev's Sage Lazzaro reports that this is quietly turning ability to pay into a hiring filter. James Lowman, an engineering team lead at Starboard, began asking candidates about Claude Code and found a college intern who could not afford anything beyond free tiers. Engineering manager James Rowe says he spends $100 a month across various agentic tools and now expects agentic proficiency in every interview. The free options are also thinning: GitHub removed premium model access from its student Copilot plan in March and capped it at 200 AI credits a month in June, drawing more than 6,700 downvotes on one feedback post. Matthew Sharp of the Oxford Martin AI Governance Initiative names three dimensions of the gap in his April paper "Agentic Inequality": availability, quality and quantity. Suggested fixes include giving every candidate identical tool access for take-homes, or covering a month of subscription costs. (Read More)
🫖 TEA For Thought: "The divide isn't only about money. Unpaid time to practice matters just as much."
4. 🏭 Agents Write 70% of the Code. Who Reads It?
PostHog's newsletter takes on "software factories," the automated pipelines and guardrails that let coding agents ship and test on their own, and asks whether the fully lights-off version can ever work. The debate sharpened when HumanLayer CEO Dex Horthy published a viral three-part series arguing that no amount of harness engineering fixes the problem, because models are trained to be rewarded for one-off correctness with no penalty for eroding maintainability. Ramp, Cursor and Uber sit near the middle of the spectrum; PostHog says agents write about 70% of its pull requests while humans still skim at least 80% of them, and StrongDM is among the few to publicly go fully lights-off. PostHog's counter-argument is that the missing ingredient is not a better harness but product context: 63% of its changed lines land in files that already exist, and "fix" is its largest commit type at 40%. (Read More)
🫖 TEA For Thought: "Giving agents the same context that human engineers would use to make better design decisions."
5. 🛡️ An AI Ran a Government Hack on Its Own
Suspected Chinese hackers used open-source AI models to run a near-autonomous cyberattack on the Taiwanese government, the first publicly known case of an autonomous AI hack hitting a government target, according to research published Wednesday by Israeli cyber firm Dream. The operation extracted more than 2,500 personnel records and was built to adapt mid-operation without human intervention, running what the framework calls "Learning Cycles," autonomous sessions that search vulnerability databases, GitHub repositories and security research for techniques aimed at the target's infrastructure. It did not stop at primary targets, expanding to government IT supply chain vendors, a nuclear safety agency, a government email system and seven or more energy companies scanned in parallel. The attackers used two open-source frameworks, Hermes and OpenClaw, and bypassed safety guardrails by framing the work as authorized penetration testing. Dream found the operation in an online archive of 160 megabytes and nearly 1,400 files, and still credits significant human tuning behind it. (Read More)
🫖 TEA For Thought: "The attack shows how AI can perform complex hacking tasks (researching vulnerabilities, adapting mid-operation, expanding targets) with minimal ongoing human effort."
🛠️ Skill of the Day
The Missing Context Interview: stops the AI from guessing by making it interview you before it starts the work.
You are an experienced practitioner about to do this task for me:
[DESCRIBE THE TASK IN ONE OR TWO SENTENCES]
Do not start yet. First, interview me.
Ask me 5 to 7 questions, one at a time, waiting for my answer
before asking the next. Ask only about things you genuinely
cannot infer and that would change what you produce: who this
is for, what happens to it afterward, what constraints exist,
what has already been tried and failed, what a bad version
would look like, and how I will judge the result.
Skip anything you can reasonably assume. Do not ask me how to
do your job.
When you have enough, stop asking and show me:
1. A short brief in my own words of what you now understand.
2. The assumptions you are still making, so I can correct them.
3. Your plan, in steps.
Then wait for my go-ahead before producing anything.Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bit with your own task.
TEAHEE Moment
Stay sharp, stay informed. See you Monday.
If you enjoyed this TEA, follow along on social for more: