Daily TEA – You Pay for AI Twice
Tencent buys Manus back, who grades the graders, Nadella’s second invoice, brain tech beyond the skull, and the AI hiring rebound
Hello, dear TEA-mates! Here is what you need to know today.
1. 🤝 Tencent Moves to Buy Manus Back From Meta
Tencent is in talks to become the largest external shareholder of Manus, the agentic AI startup, leading a repurchase that unwinds Meta’s blocked acquisition. Chinese regulators ordered Meta’s December 2025 deal reversed in April 2026 over foreign-ownership concerns. Tencent is reportedly raising around $1 billion to fund the buyback at the same $2 billion valuation Meta paid. Venture firms ZhenFund and HSG are also part of the talks. Tencent would become the single largest holder while remaining a minority overall, according to the Financial Times. (Read More)
🫖 TEA For Thought: “Sigh. I really do like Manus. If this doesn’t alarm people to the importance of democracy, I don’t know what will.”
2. 📏 Who Evaluates the Evaluations?
Google’s Data Cloud frontier AI team argues that a single benchmark score is like a pass/fail exam: it hides where an agent actually breaks. Their fix, Discovery Bench, uses an information-theory measure called surprisal to regenerate each test question at calibrated levels of ambiguity, mapping the exact point an agent fails. Testing a recall agent on Gemini 3.1 Pro, F1 scores swung from 0.81 at medium ambiguity down to 0.34 when questions turned vague, and one query that scored a perfect 1.00 at neutral phrasing collapsed to 0.00 one notch vaguer. They also found a widely trusted astronomy benchmark had broken ground-truth data. The call to action: evaluate your evals. (Read More)
🫖 TEA For Thought: “The evaluations of evaluations.”
3. 🧾 You Pay for AI Twice
In an X essay titled “The Reverse Information Paradox,” Microsoft CEO Satya Nadella argues that companies using frontier AI pay twice: once in cash, and again by leaking the proprietary know-how that makes them unique. He calls the leak “exhaust,” the prompts teams write, the tools agents call, and especially the corrections and evals they run, all of which quietly train the vendor’s model. Nadella frames it as an inversion of economist Kenneth Arrow’s 1962 information paradox, where the risk once sat with sellers and now sits with buyers. His prescription: build private learning environments inside your own “tenant boundary,” keep evals and institutional memory in-house, and avoid depending on any single model. (Read More)
🫖 TEA For Thought: “This is the very reason local models will be the future. Down the line, there might be refrigerator-like devices for people to run local models at home. The Fable-like local models are not too far away.”
4. 🧠 China’s Answer to Neuralink Skips the Surgery
Chinese neurotech firm BrainCo, founded in 2015 out of Harvard Innovation Labs and now one of Hangzhou’s “six little dragons,” is scaling brain-computer interfaces that skip surgery entirely. Its headbands and caps read electrical signals through the scalp, a sharp contrast to Neuralink’s implanted electrodes: lower risk and easier to scale, though harder to regulate. Products range from FDA-approved bionic hands that turn neural and muscle signals into finger movement for amputees to attention-tracking Focus headbands and a sleep-aid band. BrainCo has raised about 2 billion yuan (roughly $280 million), co-led by IDG Capital, and filed confidentially for a Hong Kong listing. Non-invasive devices make up around 82% of China’s domestic brain-computer interface market. (Read More)
🫖 TEA For Thought: “Sometimes you gotta think out of the box. In this situation, you just gotta think out of the skull. Apparently, China did.”
5. 💼 The AI Job Market Flips From Cuts to Creation
New Indeed Hiring Lab data suggests AI’s effect on hiring is turning from destruction to creation. US software-development job postings have grown nearly 15% since Claude Code launched in late February 2025, even as overall postings fell 7%, though software roles remain 27.5% below their pre-pandemic February 2020 level. From May 2022 to May 2026, the most AI-exposed occupations saw the steepest declines, but over the past year that relationship flipped and those same occupations rebounded the most. The recovery is concentrated: 71% of the software-posting gains came from senior roles, and 37% mentioned AI in the job title, pointing to demand for experienced people who can work with AI rather than a broad rebound. (Read More)
🫖 TEA For Thought: “Well, it’s like a gold rush. Whenever there’s gold, and AI brings almost infinite possibility, the gold diggers are needed.”
🛠️ Skill of the Day
The Knowledge Debrief: turn a finished task into a reusable memory note so the lessons don’t evaporate.
You are my debrief partner. I just finished a task and I want to capture what I learned before I forget it, so it becomes reusable knowledge instead of evaporating.
Here is what I did:
[PASTE A SHORT DESCRIPTION OF THE TASK, THE OUTCOME, AND ANYTHING THAT SURPRISED YOU]
Interview me with one focused question at a time (do not dump all questions at once). Work through: what the goal really was, what worked, what went wrong and why, what I would do differently, and any decision a future me (or a teammate) would need explained.
After the interview, output a single "memory note" with these sections: Context (2 sentences), Key Decisions and Why, What Worked, What To Avoid Next Time, and Open Questions. Keep it under 250 words, plain language, no filler. End with 3 tags I could file it under.
Paste into ChatGPT, Claude, or your tool of choice after finishing any project, and keep the note somewhere you will actually find it again.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:
Twitter/X