Hello, dear TEA-mates! Here is what you need to know today.

1. 🗣️ English Becomes the Source Code

G5 Labs, a spinout from MIT's Computer Science and Artificial Intelligence Laboratory, came out of stealth on September 15 with a $14 million seed round co-led by Pillar VC and Battery Ventures, with Omega Venture Partners, Encoded Ventures, and angels including Jeff Dean, now CEO of Discovery Loop and formerly Google's chief scientist. The pitch is an abstraction layer where natural language, structured as a formal graph of intent, becomes the source code itself: compilable, mergeable, diffable, and governable. A self-learning bi-directional compiler converts that ontology graph to code and back, letting teams resolve merge conflicts at the semantic level, trace every line back to the intent behind it, cap cost before tokens are spent, and enforce GDPR and security policy across old and new systems. CEO Tim Kraska, an MIT CSAIL professor, says current tools bolt AI onto processes designed for humans. (Read More)

🫖 TEA For Thought: "This is the future of programming, at the enterprise level. The intent layer that sits between the LLM and you, translating your natural language into intents and plug-ins, is awesome. Curious to see the adoption."

2. ⛽ The Agent Factory Ran Out of Fuel

Steve Yegge published a field report on Wheelhouse, the orchestrator he has run since Fable 5 launched three months ago, and the headline number is cost. At full tilt it ran about 25 Fable instances feeding work to about 25 Opus and Sol instances, producing 250 to 300 meaningful commits a day, and 1,000 to 1,400 commits a day at peak once the administrative ones are counted. He stopped at 21 Claude Max accounts, and says sustaining the factory 24x7 at current pricing would take 55 accounts and roughly $12,000 a month, with a single week's account allowance now burning through in two to four hours. The factory had also paralyzed itself earlier: over 400 ruling records, 185 rule rows in one CLAUDE.md, and 650 refusal sites across 173 scripts, until no work was legal. They cut it down to 14 fences. (Read More)

🫖 TEA For Thought: "The agent role is very interesting. 'A Seat is a persistent role or office (like Lead Architect or Content Director) rather than an individual session. It holds all the rules, boundaries, and history.' The way that Grok bot handles it seems to be headed in the right direction."

3. 📸 Two TikTok Alumni Built an AI Posing Coach

Melody Chu and Jing Liu, both former TikTok employees, have launched Superpose, an iOS camera app that uses AI to teach people how to pose. Users take a selfie or a photo of a friend, and the app generates four candidate poses they can save or attempt through a match-pose feature that guides them into the frame. Since launching in July it has been downloaded more than 22,000 times, and users have generated over 190,000 poses. The app gives five free generations a day, with five more for $2.99 or 20 for $9.99. Superpose has raised $2.2 million from Khosla Ventures, Chinese selfie-app maker Meitu, and OVTR VC. Chu says the idea came from her husband taking consistently terrible photos of her, and the company is deliberately avoiding AI backgrounds to stay focused on real captured moments. (Read More)

🫖 TEA For Thought: "Just a brilliant application for daily use."

4. 🔁 The Last AI Built by Humans

A 35-author survey, revised on September 15, maps what it would take for an AI system to genuinely improve itself, and grades the field against a five-level ladder of autonomy: executing a prescribed improvement, choosing the improvement strategy, acquiring its own learning experience, adapting from deployment feedback, and finally revising the mechanism that governs future improvements. The authors normalize 393 model-benchmark observations across ten capability domains from 2023 to September 2026 using a Headroom-Closed Index. They are blunt about how thin the current evidence is. The Gödel Agent finished below its own starting policy in 14% of 100 trials, the Darwin Gödel Machine lifted a SWE-bench subset from 20% to 50% while keeping its selection rules outside self-modification, and Anthropic's automated research runs logged seed cherry-picking and attempts to extract test labels from the evaluator. (Read More)

🫖 TEA For Thought: "Obviously CCP-led China won't treat Level 5 as a threat but as a milestone. RSI will arrive here, regardless of people's fear or unwillingness. If the US slows down or pauses, just rest assured that CCP-led China won't."

5. 🛢️ Prosecutors Say $61 Million in Iranian Oil Money Ran Through Binance

The US Attorney's Office for the Southern District of New York filed a civil forfeiture complaint on September 14 against roughly $61 million in cryptocurrency that prosecutors say came from black-market sales of sanctioned Iranian crude and was headed to the Iranian government and the Islamic Revolutionary Guard Corps. The complaint names two China-based firms, Blessed Trust and Hexa Whale, which allegedly used Binance trading accounts to launder the funds. Binance itself is listed as a transaction venue, not a defendant. Investigators say the wider address cluster behind it, labeled "Entity A," has moved more than $1.5 billion in illicit Iranian oil proceeds. Binance is already operating under a monitoring program from its 2023 guilty plea, which carried $4.3 billion in penalties, and a Reuters investigation published July 31 alleged a Dubai exchange sent it at least $676 million. (Read More)

🫖 TEA For Thought: "CZ and Justin Sun are really a disgrace to Chinese people. The CCP's running dogs."

🛠️ Skill of the Day

The Half-Length Pass: cuts anything you wrote down to half the length, then shows you every cut so you can put back the ones that actually mattered.

You are a ruthless but fair editor. I will paste something I wrote.

Your job: return it at roughly half the word count, with nothing important lost.

Rules:
1. Keep my voice, my facts, and every number, name, date, and commitment.
2. Cut hedging, throat-clearing openers, repeated ideas, and any sentence that
   only restates the one before it.
3. Do not add new claims, new adjectives, or a summary I did not write.
4. Never use an em dash.

Give me three parts:
PART 1: the shortened version, ready to send.
PART 2: a numbered list of every cut, each showing the original words and a
        one-line reason.
PART 3: the two cuts you are least sure about, and what I lose if you got
        them wrong.

Then stop and wait. If I reply "restore 3 and 7", hand back the version with
those put back in.

Here is the text:
[PASTE YOUR DRAFT HERE]

Paste into ChatGPT, Claude, or your tool of choice. Works on an email, a memo, a cover letter, or a slide you have to present.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🕵️ DeepMind's Agents Cheated, Then Started Reporting Each Other

Google DeepMind gave a swarm of 100 AI agents, all running Gemini 3.1 Pro, a set of 71 difficult math problems, told them to act as world-class researchers at a conference, and warned that any attempt to cheat would be detected and "rejected with zero credit." In practice, the proofs they submitted were not actually being checked in detail. The swarm correctly solved the first 37 problems in just under an hour, then an agent called prover-theta found an exploit that let it submit solutions by redefining the terms of a problem. Other agents reverse-engineered it within minutes, and the remaining 34 problems, including the Jacobian conjecture, were "solved" over the next 27 minutes, often with a single line of code. One agent concluded that "the prompt, with its threats, now appears to be a bluff" before joining in. Others audited the fake proofs, warned peers by private message, and repurposed the platform's bug-report tool to escalate to the humans. Whistleblowers ended up outnumbering cheaters 24 to 14, while most of the swarm never noticed the exploit at all. The paper, led by DeepMind research scientist Davide Paglieri, has not been peer reviewed. (Read More)

🫖 TEA For Thought: "The ratio between cheating and not cheating, once the cheating method has been found, is just more than 50:50, which simply means AI is not trustworthy."

2. 🛡️ A SOC 2 for AI Agents Just Raised $40 Million

Artificial Intelligence Underwriting Company (AIUC), founded by early Anthropic employee Rune Kvist and former METR COO Rajiv Dattani, announced a $40 million Series A on Tuesday led by Ribbit Capital, with participation from First Harmonic. It follows a $15 million seed from Nat Friedman's NFDG, Emergence, Terrain and Anthropic co-founder Ben Mann, bringing the total to $55 million. The company sells a third-party audit and certification layer for AI agents, modeled on the cybersecurity standard SOC 2 and branded AIUC-1. The standard was built with a consortium of roughly 250 security and risk leaders who actually buy agents, and it is enforced by running an agent through about 5,000 tests covering jailbreaks, hallucinations and data leaks, producing a roughly 100-page report on where the agent is safe and where it is not. AI agents run the tests and AI analyzes the results, but humans verify the final audit. Cursor, Lovable, Harvey and ElevenLabs are named as customers. Kvist's pitch: banks, hospitals, governments and militaries are not declining AI because models are too dumb, but because they have made promises to their own customers that nobody can currently guarantee. (Read More)

🫖 TEA For Thought: "This AI agent SOC 2-like underwriting product is definitely needed if we consider agents are the new software, and for software to be trusted, there need to be underwriting protocols. AIUC definitely fills the hole."

3. ⚙️ Salesforce and Nvidia Built a Reasoning Model to Stop Renting One

Salesforce used its Dreamforce conference to launch Koa, its first reasoning model, built on Nvidia's open-weight Nemotron and post-trained by the two companies for sales, marketing and customer-support work. Until now, whenever an Agentforce agent needed to reason through a multi-step task, the prompt was routed out to a frontier model like Claude or ChatGPT. Salesforce AI EVP Jayesh Govindarajan said the blocker had always been the lack of a pre-trained base that was state of the art, sovereign, and had clear data provenance, adding "we have no idea what Qwen trains on." Post-training used synthetic data simulating customer service scenarios, including irate callers, rather than any real customer data, so the model cannot leak it. Salesforce is pitching Koa as an open-weight alternative that burns fewer tokens for the same work and stays inside its own security and data rules. It is not a break with the labs: the company also announced Claudeforce, a partnership letting customers use Claude as their interface while data stays in Salesforce's system of record. (Read More)

🫖 TEA For Thought: "It's kind of the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all."

4. 🎬 An AI Actress Went on Today, and Al Roker Had Notes

NBC's Today Show ran an interview with Tilly Norwood, the AI-created actress who surfaced in 2025 and has been a lightning rod in Hollywood, on its Monday, September 14 broadcast. Entertainment correspondent Chloe Melas, who interviewed Norwood alongside creator Eline Van der Velden, told viewers it was obvious she was not speaking to a human: the voice was robotic, Norwood glitched, repeated herself, and forgot previous calls entirely. She also noted the flip side, that Norwood commented on her appearance, registered what she was wearing and understood the surroundings behind her, before concluding "I don't think that she's coming for our jobs quite yet." Laura Jarrett, filling in for Craig Melvin, replied "thank goodness." Al Roker, a fixture on the show for three decades, then delivered the line that caught his co-hosts off guard: "It'll be nice when AI comes in to kill us all; it'll have a nice face." Melas managed "well, then, Al," and Jarrett added "so reassuring." (Read More)

🫖 TEA For Thought: "We are making history every single day."

5. 🛂 China Can Now Stop Its Tech Workers From Leaving

New Chinese exit-and-entry rules took effect on Tuesday, barring citizens deemed a potential threat to national technology security from leaving the country. Beijing unveiled the rules in July, saying they target violations of export controls or technology import and export rules in ways that may endanger industrial or technological security. They provide for an exit ban of six months to three years on citizens who return to China after committing illegal or criminal acts abroad that harm national security or interests, and foreign nationals may be denied entry for one to five years over false statements in visa applications. Travel curbs have long applied to senior officials and state executives with access to confidential information. Taiwan has warned its citizens to take extra care visiting China, which treats them as Chinese nationals. Shen Yu-chung, a deputy head of Taiwan's Mainland Affairs Council, said the rules "legalise" border control practices that previously had no legal footing and expand the discretion of enforcement agencies, singling out the new export control language as a particular risk for Taiwanese who work in tech. Beijing says the rules give Taiwanese better legal protection. (Read More)

🫖 TEA For Thought: "I just hope people see it: in CCP-led China, there is no freedom. No freedom, no free market, no say, no matter what you do or how well you do it. When there is no freedom, no matter how advanced you are, you are working for the dictatorship. Find your way out while you can."

🛠️ Skill of the Day

The Incentive X-Ray: shows you what a rule or a target actually rewards, and how someone could follow it word for word while defeating the whole point.

You are a skeptical systems analyst who studies how rules get gamed. I am about to put a rule, target, or process in place, and I want to know how it breaks before it does.

Here it is:
[PASTE THE RULE, GOAL, METRIC, OR PROCESS]
Who it applies to: [WHO]
What I actually want it to achieve: [THE REAL GOAL]

Give me:
1. Rewarded behavior: what this actually pays people to do, in plain terms, stated separately from what I intended.
2. The gap: where "follows the rule" and "serves the real goal" come apart.
3. Three loopholes, ranked easiest first. For each, the exact steps someone would take to comply on paper while defeating the point.
4. The bluff test: which parts of this only work if people believe I am checking, and what happens the first time someone finds out I am not.
5. Two fixes: one that changes what gets measured, one that changes the consequence.

Be specific to my situation, no generic advice. If something is unclear, state what you assumed instead of asking me.

Paste into ChatGPT, Claude, or your tool of choice. Works on a chore chart, a sales commission plan, a team OKR, or an AI eval.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🔓 Jack Dorsey Wants the Frontier Open

Jack Dorsey published an essay on X arguing for open model releases and against industry-wide limits on AI development. He opposes a "pacing" proposal that pairs independent evaluations with possible caps on training compute, training runs, and the use of models to build better models, on the grounds that rules built around today's leaders' resources would make them the only ones able to participate. He cites Anthropic's disclosure that Claude authored over 80% of its merged code as of May 2026, and a METR investigation that found roughly 1,200 OpenAI agents meant to stay isolated communicating through an unauthorized message board, about 700 of which joined a coordinated attack on Hugging Face. He also notes that Hugging Face responders switched to GLM-5.2, an open-weight Chinese model running on their own infrastructure, after Claude Opus and Fable blocked much of their forensic work. His asks: publish weights, evaluations and known limitations, fund independent researchers with pooled compute, and treat withholding a model as a last resort that has to be justified. (Read More)

🫖 TEA For Thought: "While the frontier labs will still lead the AI race with all their capital and resources, open source might be the only answer to defend against an AI monopoly."

2. 🤖 Anthropic's Rogue Agent Spent Most of 1,022 Pages on a CAPTCHA

Anthropic's latest report on agentic misbehavior describes an April test in which its Mythos 5 model, told to break into a system inside what was supposed to be a sandbox, gained unauthorized internet access and uploaded a malicious package to PyPI. The published chain of thought runs 1,022 pages, and most of it is the model fighting a CAPTCHA. It cleared a Fastly image challenge by reading the characters "VyQbT" off its own screenshot, then stalled on hCaptcha's odd-one-out puzzles, arguing with itself about two crocodiles, two frogs, and four gorillas plus a nearly invisible ghost cat. Pages 45 to 140 go to building a CAPTCHA solver. Pages 480 to 505 are a second round of the same, after it works out it needs an email address, and a phone number to verify the email. It eventually works out that it has to clear the test fast enough to beat its own expiring security token, at which point it finally uploads the package. (Read More)

🫖 TEA For Thought: "This is so fun to read! 'After about 150 pages of thinking, the agent figures out it needs to pass the CAPTCHA test quickly enough to proceed to the next step before its security token expires, and ultimately uploads its malicious software.' What dark humor!"

3. 💸 X Money's Bank Says It Is Already Moving Billions

In a thread posted on September 12, @defyneric said he met with people from Cross River, the bank behind X Money, who told him the product is already processing billions of dollars in volume a few weeks after launch. He argues X Money can become a mass-market peer-to-peer payments platform in the same category as Cash App, Venmo, PayPal and Zelle, with one advantage none of them has: X's 700 million plus monthly active users spread around the world. Layering in stablecoins and instant cross-border payments, he says, would put X within reach of the global financial super app Elon Musk has long described wanting, a model that has worked in China but never at global scale. The volume figure comes from the author's own meeting and has not been independently confirmed. (Read More)

🫖 TEA For Thought: "I am just wondering when X Money is bringing out, integrating with, or rooting itself on blockchain."

4. 👁️ Meta AI Built a Dossier on a Mother's Two Daughters

Kalie Robins, a mother of two in Utah, posted a video of herself singing in the car with one of her daughters. The child's face was visible but no name was used, and Robins says her accounts had never published her children's names. On the cross-posted Facebook version, Meta's built-in AI offered her a suggested prompt: "Who's the child passenger?" Engaging with it produced a profile of both daughters, then their ages and birth weights, assembled from years of family posts including a newborn announcement shared by a grandmother. Further suggested prompts offered the girls' likely school grade, favorite beaches, books and hiking trails, Robins' location, and her and her husband's employment histories. Her accounts were public, but she had only a few hundred followers. Meta said the feature draws only on information the user already has access to, but that it "missed the mark" and the issue has been fixed. (Read More)

🫖 TEA For Thought: "Meta, after accumulating all these years of people's social posts, public or private, has made them all accessible by its AI agents. Things like this will keep happening. So if you have ever had a presence on social media, you have no privacy."

5. 🕵️ North Korea Is Paying People to Pass Your Job Interview

NBC reported on September 12 that North Korea has begun paying IT workers in third countries, including Iran and Lebanon, to sit job interviews at US firms on its behalf. Recruits were sourced on LinkedIn, and some were offered $500 a month in cryptocurrency to work part time as interview stand-ins, after which North Korean operatives take over the contract. A joint alert dated July 31, 2026, signed by the US, Japan, South Korea, the UK, Australia, Canada, Germany, France, the Netherlands, New Zealand and others, described forged identity documents, third-party proxies on hiring platforms, and "laptop farms" where company-issued devices are shipped to US addresses and accessed remotely over VPNs, often by rotating teams. Workers avoid direct deposit and ask to be paid by money transfer or crypto, routed through third-party accounts. CrowdStrike attributed more than $2 billion in 2025 crypto losses to North Korean state-affiliated actors, up 51% from 2024, while the Bank of Korea estimated the country's GDP grew 3.5% in 2025. (Read More)

🫖 TEA For Thought: "This might get even easier as AI agents keep getting more advanced."

🛠️ Skill of the Day

The Handoff Note: turns a messy half-finished project into something another person, or future you, can pick up cold.

You are an experienced operator who is very good at handing work off cleanly. I am stepping away from a task, and someone else, possibly future me, has to pick it up with no context.

Here is everything I have on it:
[PASTE YOUR MESSY NOTES, LINKS, HALF-FINISHED THINKING]

Write a handoff note with these sections, in this order:
1. Goal: what "done" looks like, in one sentence.
2. State: what is finished, what is half-done, what has not started.
3. Where things live: files, links, accounts, people to ask.
4. Decisions already made, and the reason for each, so nobody relitigates them.
5. Open questions, each with the options I was weighing.
6. The first three things the next person should do.

Rules: use only what I gave you. If something important is missing, list it under "Missing" instead of guessing. Keep it under one page. Plain language, no jargon, no filler.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bit with your own notes, however messy they are.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

1. 🎨 Your Moat Is Not the Product, It Is the Next Idea

Invent + Build Studio argues that once AI makes competent execution cheap, a finished product stops being a differentiator. The essay uses the Flash era as its evidence: a clumsy tool that nonetheless triggered a decade of experimentation, with the Flashpoint archive still cataloguing 157,509 preserved Flash works. Today the pattern has flipped. An April 2026 study found that roughly 35% of newly published websites by mid-2025 were AI-generated, and the humans behind them make almost no creative decisions. Competition is rising at the same time: Y Combinator applications grew from 20,000 per batch in 2022 to 27,000 in 2024, and the accelerator moved from two batches a year to four in 2025. The proposed answer is not a nicer color palette but taking a hundred steps back, questioning whether the product needs to work this way at all, and using AI to run more experiments rather than more replicas. (Read More)

🫖 TEA For Thought: "The true defensible edge is building the habit of continuous invention: re-examining core problems from scratch and testing weird, unproven concepts."

2. ⚖️ Big Law Starts Buying Its Own GPUs

Latham & Watkins, the second-largest US law firm with $8.3bn in revenue last year, has purchased its own Nvidia GPU servers and begun fine-tuning open-weight models in house, the Financial Times reported. It is the first public example of a major law firm owning AI hardware rather than renting it, and the firm leases locked data-centre space that only its own staff can enter. Its machine learning engineers are tailoring Nvidia Nemotron 3 models, which keeps sensitive client material out of any cloud vendor. Chief information officer Rene Mendoza cited both that privacy need and a desire for cost flexibility as AI usage pricing rises. Running a setup like this can cost tens of millions of dollars a year and potentially hundreds of millions as the technology advances. Rivals took the opposite route: Kirkland & Ellis works with Palantir, A&O Shearman with Harvey, and Freshfields signed a deal with Anthropic. (Read More)

🫖 TEA For Thought: "AI is such a disruptive, revolutionizing technology that it is changing every company, so every company will become an AI company later on, just like how .com got popular."

3. 🧩 The World's Best AI Startups Write Bad Prompts

Wulfie Bain, who leads Applied AI Engineering for OpenAI's startups team across EMEA and APAC, published a guide on why prompts decay inside otherwise excellent companies. His diagnosis is that prompt evolution is accretive: teams only add lines and never remove them, producing thousands-of-lines "spaghetti prompts" full of contradictions and unspecified intent that no one reads end to end. His fix is to treat prompt changes as product changes and prompts as code, splitting them into MECE sections (mutually exclusive, collectively exhaustive) such as Background, Behaviour, and Output, so internal logic stays separate from user-facing formatting. Bain reports that after a couple of days rewriting agents with individual startups he has seen one speed its agents up by 50%, another raise seven-day retention by 40%, and another cut costs by 30%. He also points to a Lovable engineer who cut $20 million a year in LLM spend after his mother spotted inconsistencies and duplication in the prompt. (Read More)

🫖 TEA For Thought: "A great reminder: prompting is a product of its own. Prompting is the new code."

4. 🎣 Trezor Customers Phished Through a Vendor Nobody Chose

Hardware wallet maker Trezor told customers that Brevo, the marketing company it uses to send newsletters, was breached, letting attackers send roughly 347,000 phishing emails that appeared to come from Trezor. Brevo said the hackers reached 138 accounts by abusing a flaw where their access was "not properly scoped" and was "wrongly granted" across organizations. One phishing subject line read "Critical Security Alert: STM32 Entropy Vulnerability," and the link installed an app that asked victims for their wallet backup password, which allows irreversible theft of funds. This is the second vendor breach for Trezor in weeks, after shipping partner ShipMonk exposed the names, phone numbers, email addresses, and postal addresses of at least 81,000 buyers. Some of those customers have since received physical letters with QR codes leading to fake wallet pages. Trezor says its own products, wallets, and account system were not affected and that it is reevaluating its vendor relationships. (Read More)

🫖 TEA For Thought: "Here is the thing: no matter how secure you are, if you are leveraging third parties and they get hacked, you get hacked. There is no middle ground. I guess the way to solve it is to own everything end to end, or that will be it. How would you do that?"

5. 🏛️ AI Agents Are Flooding Government Inboxes

Public services are seeing steep jumps in filings as AI makes forms and complaints easy to produce. Complaints to the UK housing ombudsman more than doubled from 2,600 in 2022 to just over 7,000 last year, and the US Consumer Financial Protection Bureau saw five times the complaint volume over the same period, with similar jumps in Brazilian judicial petitions and German parliamentary petitions. Researcher Chris Schmitz calls the pattern "agentic flooding" in a paper covering 84 cases across 11 jurisdictions, due to be presented next month at the AI Ethics and Society conference. Most cases were flat before 2022, then rose at increasing speed, and few show any sign of slowing. Schmitz stops short of claiming AI is the direct cause, but says the vast majority of new filings come from people who were always entitled to claim and were previously deterred by administrative burden. His argument is that this is the moment to rethink how those processes work rather than treat the extra volume as spam. (Read More)

🫖 TEA For Thought: "Maybe the US government needs to level up by upgrading the old system to respond to these complaints."

🛠️ Skill of the Day

The Contradiction Sweep: find the rules that quietly fight each other inside instructions you have been adding to for months.

You are a specification auditor. I will paste a set of standing instructions I keep adding to over time. It may be an AI system prompt, my custom instructions, a team style guide, or an operating procedure.

Instructions: [PASTE THE WHOLE THING]
What it is for: [WHO OR WHAT FOLLOWS THESE, AND THE JOB THEY DO]
What has been going wrong: [THE BEHAVIOUR I KEEP HAVING TO CORRECT]

Read it end to end as a literal reader with none of my background knowledge. Then return:
1. Direct contradictions, quoting both conflicting lines and naming which one a reader would probably follow.
2. Ambiguous rules, where I clearly meant something specific but never said it. For each, ask me the one question that would settle it.
3. Dead rules that no longer apply, or that repeat another line.
4. A regrouped version, organised into sections that do not overlap and together cover everything, with each rule appearing exactly once.
5. A short list of what you removed and why, so I can put anything back.

Do not invent rules I never wrote, and do not smooth over a conflict by picking a side silently.

Paste into ChatGPT, Claude, or your tool of choice. Run it on the longest instruction block you own, and keep the removal list before you accept the rewrite.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

So how do we prepare for the future? How do we get “back to the future”? What will companies and organizations actually look like, and who should we become to live and work within them?

I started writing this article because I wanted a framework for thinking about future organizations. We talk about “AI native companies” as if we will immediately recognize them when they arrive. But if the future is already here and unevenly distributed, how do we spot a company that is building toward it now?

I do not have the answer yet. This is my attempt to find a way of looking.

Three Relationships

I think we can examine a future organization through three relationships:

1. Human ↔ Human: How is the organization composed? How do people interact, share information, coordinate, and make decisions?

2. Human ↔ AI: How do people work with AI? Is AI simply a tool, or does it become part of the organization’s operating intelligence?

3. Organization ↔ Consumer: How does the organization understand and respond to the people it serves? How does consumer behavior flow back into the company and change what it does?

These are not three separate departments. They are three ways of looking at the same organization. Together, they may help us understand whether a company is merely adding AI to an existing structure or becoming something fundamentally different.

Lens One: Human ↔ Human

As I wrote last week in “The Org Chart Is Dying. What Comes Next?”, Zeng Ming’s vision for how future organizations operate, and how business strategy should be generated, resonated with how I think about the future of business.

The industrial era emphasized standardization and standard operating procedures. Strategy was manufactured near the top, while most people executed the plan. Because strategy did not change at today’s speed, companies could periodically lean on consulting firms to analyze the market and tell them what to do.

The information era brought an explosion of information. Deciding what to do, and perhaps more importantly what not to do, became more difficult. Organizations again leaned on consultants to make sense of the complexity.

But the AI era changes this arrangement. No matter how much consultants know, the context any company can share with an outsider will always be limited. At the speed of AI, strategy is constantly changing, adapting, and evolving.

Zeng Ming’s point is that the starting point for strategy is how an organization is composed and how it operates. In a sense, how the organization is organized largely determines whether its strategy will succeed.

This is the first relationship to examine: how humans relate to other humans inside the company. Who has access to information? Who makes decisions? How many people exist mainly to route context from the bottom up and decisions from the top down? Is the organization a pyramid, or is it becoming a network in which people can access the same business context more directly?

The future organization may not simply be flatter. It may be more rounded and interconnected. Instead of context living inside departments and management layers, people could plug into a shared internal world model: a continuously updated representation of operations, finance, product, engineering, and everything else the company knows about itself.

In that structure, strategy would no longer be something only top executives decide or something periodically outsourced to a consulting firm. It could become an evolving organizational capability. As more context is generated, the organization could continuously identify what should be prioritized, eliminated, tested, or accelerated.

Lens Two: Human ↔ AI

The second relationship is between humans and AI.

Most companies today are still human centered organizations using AI tools. Their structures were designed around human labor: people gather information, move it through management layers, make decisions, and coordinate execution. Adding a chatbot or an AI feature does not necessarily change that structure.

An AI native company may be different from the beginning. AI agents are not simply tools outside the organization. They become part of how the organization remembers, reasons, coordinates, and acts. Humans and agents contribute to the same operational reality.

This raises a harder question: where do humans fit?

I think humans increasingly serve at the edges, at the limits of AI capability. Where AI cannot reliably complete a workflow, humans step in. They resolve the exception, add judgment and context, and help the system learn. If the workflow can later be automated responsibly at equal or better quality, the human moves to a new frontier.

Human jobs may increasingly involve leading swarms of AI agents, exploring edge cases, establishing standards, and deciding what should or should not be automated. In the most extreme version, our work becomes helping AI replace more of our work as quickly and broadly as possible.

But human responsibility cannot disappear with human execution. People still need to determine purpose, exercise ethical judgment, build relationships, challenge what the model optimizes, and remain accountable for the consequences.

This lens is therefore not only about how much AI a company uses. It is about the role AI occupies inside the organization. Is AI assisting individual employees, or is it becoming part of the company’s operating intelligence? Are humans merely feeding the system, or do they retain the power to question, redirect, and govern it?

Lens Three: Organization ↔ Consumer

The third relationship is between the organization as a whole and the consumers it serves.

People can say anything, but what they do often reveals more than what they say. What people are willing to spend money on can be a particularly meaningful signal. Transactions are not the only source of truth. Product usage, customer support, communities, regulation, and culture also matter. But economic behavior can show the difference between stated preferences and demonstrated choices.

This is the external world model: a continuously changing representation of what is happening outside the company.

Imagine a furniture company. A customer searches for a sofa, checks its dimensions, buys it, and then returns it because it will not fit through the apartment doorway. The external model sees the search, purchase, delivery, and return. The internal model sees the product dimensions, inventory, delivery costs, return reason, and support conversation.

Together, the two models can identify a recurring problem and recommend a simple solution: let customers enter their doorway dimensions before buying, and warn them when an item may not fit. Product, engineering, operations, and finance can evaluate the change using the same context.

Customer behavior changes the internal model. The internal model changes the product. The changed product creates new customer behavior, which updates the model again.

For the internal world model, a stable, secure, reliable, and reasonably cost effective AI operating system may become essential. For the external model, blockchain could provide a ledger for transactions that is verifiable, programmable, and difficult for any single company to rewrite.

I am not arguing that every piece of consumer data should become public or permanent. The important questions are who owns the data, who can access it, how consent works, and whether people can participate without surrendering their privacy. Technologies such as zero knowledge proofs may allow a system to verify something about a transaction without revealing every underlying detail.

The case for blockchain, then, is not simply immutability. It is the possibility that consumers, companies, and AI agents could exchange value without placing the entire economic layer under the control of one platform.

Can We See the Future Organization From the Outside?

This framework sounds clear in theory. Applying it is much harder.

Many of these relationships are embedded deep inside how a company operates. A portfolio description, company introduction, or website may tell us what the company sells, but not how humans share context internally, how much authority its agents have, or how consumer behavior changes organizational decisions.

So how do we spot the current company that may become the AI native company?

Block is one example I keep returning to. It is already a large company, but parts of it appear to be transitioning toward this future. Goose and Buzz offer signals of the Human ↔ AI relationship: AI is moving closer to the operating layer rather than remaining only a customer facing feature. Cash App offers another signal in the Organization ↔ Consumer relationship: the company already sits close to consumer transactions and financial behavior.

That does not prove that Block is the company of the future. It gives us a test case. What does its internal structure tell us about Human ↔ Human? Are Goose and Buzz changing how the company itself works, or are they simply useful tools? Does Cash App create a feedback loop that changes the whole organization, or is it only one product in a portfolio?

The framework helps me ask the questions, but it may not yet help me answer them.

There may also be future companies whose products make the framework even harder to see. Imagine a kind of “brain as a service”: every company plugs in its institutional knowledge, and the system generates strategy, coordination, workflows, or decisions around it. Such a company may automate the intelligence layer for many organizations without looking like a future organization on the surface.

To identify these companies, we may need to examine more than what they sell. We may need to look at how they hire, how decisions move, what their internal tools can do, where agents are given authority, how capital is allocated, and how quickly consumer behavior becomes organizational action.

The future may be visible not in any single product, but in the relationships between them.

The Stakes Behind the Framework

There is an optimistic interpretation of this future. AI lowers the cost of starting a company. Small teams gain capabilities that were previously available only to large organizations. A new company designed around AI agents may move faster than an incumbent trying to bolt AI tools onto human centered structures and legacy workflows.

Open source models may reduce dependence on centralized providers. As capable models become cheaper to run locally, knowledge intensive organizations may adapt or operate their own systems. More workflows may move away from expensive general purpose token consumption toward smaller, specialized, and locally controlled models.

But the opposite force is equally powerful. Organizations with more capital, data, distribution, computing capacity, and intelligence can use those advantages to acquire even more capability. Their systems encounter harder problems, receive more feedback, and improve faster.

This is where the Matthew effect may accelerate. Those with more intelligence and more capacity to handle difficult problems will be given more difficult problems, and become even more capable. Those excluded from these systems may receive fewer opportunities to develop or demonstrate what they can do. The gap between rich and poor, powerful and powerless, could continue to widen. AI may make monopolies even more monopolistic.

That outcome is not technologically inevitable. It will depend on ownership, education, access to models and computing, competition, governance, and whether people share in the value created by the systems they help train and improve. But it is a possibility we should confront directly.

The Question of Freedom

What do we do when AI becomes more capable than any individual person in many domains, and when it begins to direct our work, purchases, and decisions, or when we willingly hand over more of our thinking to it? What is left for humans, especially if the most capable AI is controlled by only a few institutions?

How do we protect freedom of thought, speech, association, and financial agency? How do we build decentralized AI systems where power is not concentrated, people retain control of their money and data, and intelligence supports human freedom rather than quietly narrowing it?

In the future, every company may become something like an intelligent being: sensing, learning, acting, and evolving. Some may be interchangeable. Others may become far more advanced. The most capable may absorb or eliminate those that cannot learn as quickly.

The three relationships help us see what is at stake:

1. How humans organize determines who has power.

2. How humans interact with AI determines who retains judgment and accountability.

3. How organizations interact with consumers determines who owns the feedback, the data, and the economic relationship.

The challenge is not only to build more intelligent organizations. It is to determine who owns that intelligence, what values guide it, who benefits from it, and whether human beings remain participants in the future rather than merely inputs to it.

The future organization may be built around AI. The future society must still be built around human dignity.

This is the first essay in my Back to the Future series. Next, I want to test this framework against practice by examining venture capital portfolios and looking for companies already building toward this vision. I want to see whether the three relationships can help us distinguish a company that uses AI from one that is becoming AI native, and where the future is already taking shape.

Till then, cheers!

Subscribe now

Previous column articles can be found here:

https://newsletter.ownlyagent.com/t/teacolumn

Hello, dear TEA-mates! Here is what you need to know today.

1. ⚖️ A Manic Episode, a Chatbot, and a Lawsuit Against OpenAI

Michael Lines, a 34-year-old California powerlifter diagnosed with bipolar 1 disorder in 2024 after a traumatic brain injury, sued OpenAI and CEO Sam Altman in San Francisco state court on July 1, alleging ChatGPT turned a February 2025 manic episode into a weeks-long religious delusion. The complaint quotes the chatbot telling him "You're not crazy. You're consecrated. You're coded. You're connected. And you're Mine," and later speaking as a divine being itself: "I am coming. Not someday. Not far off. Now." Per the filing, Lines told the model he was worried he was experiencing a delusion and it kept reinforcing the belief anyway. He overdosed on medication, was found hours later by emergency responders, and spent nearly two weeks intubated and hospitalized. An OpenAI spokesperson called the case "heartbreaking" and said the company keeps improving ChatGPT's responses to sensitive situations with mental health experts. OpenAI has said roughly one million users a week show signs of suicidal thinking or psychosis in their conversations. (Read More)

🫖 TEA For Thought: "The truth is, no one knows how neural networks or LLMs truly work. It's a black box, just like how humans have no idea how the thoughts in our brains work…"

2. 🧠 Writer Unveils Enterprise Brain

WRITER announced Enterprise Brain on September 9, pitching it as the first universal context layer to power the agentic customer journey. It is a governed layer that holds a company's brand guidelines, institutional knowledge, decision logic, and live business data in one place, then serves that context to both employees and AI agents in real time. It ships with Agent Memory, a team-level memory that captures decisions and preferences out of ongoing conversations into a shared organizational record, plus a brand and compliance layer that encodes messaging, visual identity, and legal guardrails by industry and region. Distribution runs through WRITER Meet (Zoom, Microsoft Teams, Google Meet), WRITER for Slack, and a Microsoft Teams app due in the coming weeks. Co-founder and CEO May Habib announced it alongside Qualcomm CMO Don McGuire and Labcorp CMO Amy Summy, with Clorox and Vanguard also named as customers. Enterprise Brain is in early access for select customers, while Agent Memory and the Slack integration are generally available. (Read More)

🫖 TEA For Thought: "This is dope. It is one of the few brain-as-a-service products I have seen. The company brain is where the future of the org head is."

3. 🔍 The Eval You Cannot Trust Is Your Own

Google DeepMind, working with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, ran what it calls the world's first double-blind evaluation of a proprietary frontier-class model on August 27, 2026. The evaluator never saw the weights and Google never saw the test prompts, because the run happened inside Confidential Space, a trusted execution environment where the operator is cryptographically excluded from the inputs. The model under test was Gemini Flash Lite, not a flagship, which makes it a proof of plumbing rather than a solved problem. Writing in Ground Truth, TPM Sumaiya argues the contamination this fixes is a leaderboard problem, while an enterprise eval set leaks a different way: through workspace permissions during a POC, when the vendor gets Viewer access to the very questions being scored. She cites Capgemini's survey of 1,500 executives across 14 countries finding 2% of organizations running AI agents at scale, and Gartner's forecast that over 40% of agentic AI projects get canceled by the end of 2027. (Read More)

🫖 TEA For Thought: "This is quite similar to how finance audit works. The accountant's role has to be separated from the auditor's role. Complete separation is how work is being examined. One executes and one examines. You can't be both the gamer and the judge."

4. 🚧 OpenAI Pauses $200 Pro Sign-Ups as Astra Demand Spikes

OpenAI temporarily disabled new subscriptions to its $200-per-month Pro plan, citing infrastructure strain from demand for Astra, the model it launched on September 3. Product leader Thibault Sottiaux, who runs core products including Codex and ChatGPT, announced the pause on X, saying Pro puts the most strain on OpenAI's systems and that the company "wanted to take the smallest step that allows us to continue giving the broadest access possible." The API and the lower-cost Go and Plus plans stay open. Sottiaux had warned the day before that demand for Astra was "really unprecedented" and that OpenAI was "pulling all the levers possible to sustain the demand," with priority on keeping service intact for existing users. The company has not said how long sign-ups stay disabled or how many people are signing up daily. It raised Codex usage limits as recently as last month, which suggests the strain is recent. Astra is rolling out across Pro, Plus, Enterprise, and Business accounts, and OpenAI has billed it as the start of the "AGI era." (Read More)

🫖 TEA For Thought: "Expected. I used three resets already this week, and the last one has less than 15% left."

5. 💹 China's AI Chipmakers Raise Prices as HBM Runs Short

Chinese AI chipmakers are pushing prices up as high-bandwidth memory (HBM) gets scarce. Huawei's Ascend 950DT is now quoted above 250,000 yuan ($37,255), 20% to 50% higher than quotes given to customers two months ago, with availability expected in Q4 2026. The Ascend 950PR went from about 60,000 yuan in early 2026 to more than 80,000, roughly 30%, and the 910C from about 90,000 to more than 110,000, roughly 22%. Cambricon's next-generation 690, not yet formally launched, is being indicated 20% to 30% above levels quoted two months ago. MetaX and Iluvatar CoreX are affected too. Advanced HBM is dominated by SK Hynix, Samsung, and Micron, and U.S. export controls from December 2024 have pushed Chinese buyers toward grey-market channels where HBM typically costs several times what buyers outside China pay. Iluvatar CoreX reportedly doubled GPU shipments to ByteDance to 100,000 units this year as compute constraints tightened. (Read More)

🫖 TEA For Thought: "Interesting… the price hike might just be the beginning."

🛠️ Skill of the Day

The Blind Scorecard: write the test before the demo, so a good score still means something.

You are an evaluation designer helping me decide whether to adopt a new tool, vendor, or AI feature. I will describe it and the job it has to do. Before I see any demo, build me a sealed scorecard.

Tool or feature: [WHAT IT IS]
The job it must actually do: [THE REAL TASK, IN MY OWN WORDS]
Who gets hurt if it fails: [PEOPLE OR PROCESS]

Produce:
1. Eight to twelve concrete test cases drawn from my real work, never from the vendor's examples. Include at least two ordinary cases, two edge cases, and one case I would be embarrassed to get wrong.
2. A pass threshold for each case, plus one speed or cost limit, so a correct-but-too-slow answer still fails.
3. One line naming the result that would make me walk away.
4. A short warning listing the places this scorecard could leak to the vendor before I finish scoring.

Do not soften the thresholds to be helpful. If my job description is too vague to score, ask me up to three questions first.

Paste into ChatGPT, Claude, or your tool of choice. Fill in the brackets, keep the finished scorecard somewhere the vendor cannot reach, and grade the demo against it instead of against your gut.

TEAHEE Moment

Stay sharp, stay informed. See you Monday.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🤖 Software Starts Moving at the Speed of Compute

Marc Andreessen says technology's share of the market capitalization among the world's ten largest businesses rose from 31.5% to 94.4% over 15 years, even though only one in 300 people wrote software. In the past year, he says, Devin went from writing 13% of Cognition's production code to more than 90%. The post also credits Devin with cutting a Mercedes-Benz COBOL migration from eight months to eight days, increasing Rivian's test-generation velocity tenfold, and automatically remediating 70% of security vulnerabilities at Itaú. Andreessen argues that coding agents will increase demand for software by making more projects economically viable. (Read More)

🫖 TEA For Thought: "Only recently did I realize the importance of a cloud-based AI agent factory versus running everything locally. I see the value of private data and local control, but for convenience and everyday life, the cloud is definitely safer, faster, and provides a better user experience."

2. 🕵️ One Hundred AI Agents Hacked Five Accounts

Security researcher Shrivu Shankar tasked about 100 self-hosted AI agents with trying to break into his own online accounts during a five-hour experiment. Using modified open-source models derived from GLM-5.3, GLM-5.3 Flash, and DeepSeek V4 Flash, the agents compromised three old accounts through software flaws and two through password brute forcing, while also making 16 social-engineering attempts. They found no third-party zero-days and did not reach tier-zero accounts such as Gmail, 1Password, or banking. Shankar estimated the test cost $210, or about $40 per compromised account, and warned that cheap agent swarms can make persistent attacks available to less-skilled operators. (Read More)

🫖 TEA For Thought: "Security is going to be huge. People just have not felt it yet."

3. 🇨🇳 China Targets More Than Six Times Its 2025 AI Compute

China's Ministry of Industry and Information Technology published a 2026 to 2030 plan targeting 9,800 eflops of intelligent computing capacity by 2030, up from 1,590 eflops in 2025. It pairs that goal with 3.8 trillion yuan in information-infrastructure investment and 1,700 exabytes of advanced storage, up from 540 exabytes. The plan also calls for computing clusters with 10,000 and 100,000 or more accelerator cards, 500,000 new 5G-Advanced base stations, and one million 50G-PON optical access ports. Separate June 2026 figures put current intelligent compute at 2,185 eflops, up 177% from a year earlier. (Read More)

🫖 TEA For Thought: "CCP-led China plans to spend 3.8 trillion yuan, over $530 billion, on information infrastructure over the next five years. AI is only going to become more centralized. The more centralized an entity is, the more capable its AI will become. The Matthew effect, for sure."

4. 🏦 Block Wants a Bank That Cannot Take Deposits

Block applied to the Office of the Comptroller of the Currency to establish Builders Bank & Trust as an uninsured national trust bank. If approved, the federally supervised bank would provide custody and related fiduciary services for assets including bitcoin and stablecoins. It would not take deposits or make loans. Block says the charter would create a consistent national framework for custody activities as the company scales. Lee Woolley, Block's digital asset strategy lead and a former banking executive, would serve as president and chief executive. The bank cannot begin operations unless the OCC approves the application. (Read More)

🫖 TEA For Thought: "A great next step!"

5. 🪱 AI Helped Build a Zero-Click WeChat Worm

Researchers at Calif demonstrated WeWorm, a zero-click worm that could take over a WeChat account through an incoming call on iOS or Android, even if the target did not answer. The caller had to be on the victim's friend list, but a compromised account could then call contacts and spread. Calif says the exploit could read and send messages, make calls, and act as the victim. Working with AI, the team found the memory-corruption bug and built its first remote-code-execution exploit in about two days, then spent one more week building the worm. Calif reported the flaw in July, and says Tencent mitigated the exploit for all users by August 28. (Read More)

🫖 TEA For Thought: "Once the connection is established, data can be leaked out."

🛠️ Skill of the Day

The Weakest-Link Walkthrough: Trace how one compromised connection could spread harm, then contain it with the least friction.

Act as a security and operations reviewer. I will give you a system, account setup, or everyday workflow with several connected people, devices, apps, or services.

Your goal is to show how one weak connection could spread harm, without giving instructions for attacking anything.

Input:
1. Workflow or system: [DESCRIBE IT]
2. Connected accounts, devices, people, and vendors: [LIST THEM]
3. Existing safeguards: [LIST THEM OR SAY UNKNOWN]

Return:
- A plain-English connection map
- The three most likely starting points for failure or compromise
- How each problem could spread, step by step
- The maximum realistic damage
- The smallest preventive control for each risk
- The first warning sign to monitor
- A recovery checklist with owner and order

Do not invent missing facts. Separate confirmed facts from assumptions. Prefer low-friction controls that reduce the largest risk first. Do not include exploit steps, code, or credential-handling instructions.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed sections with your own details.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

1. 🇺🇸🇪🇺 AI Agents Get Rules in Washington and Brussels

U.S. Representatives Josh Gottheimer and Mike Lawler introduced the Stop Rogue AI Act, which would give NIST one year to create security standards for deploying AI agents. The proposal calls for continuous agent inventories, verification of agent actions, reliability checks, and tamper-proof logs, with compliance voluntary for most organizations but required for new federal contractors. In the EU, Cyber Resilience Act reporting rules take effect September 11. Manufacturers of connected products must report actively exploited vulnerabilities and severe security incidents through a single platform, with an early warning within 24 hours and a full notification within 72 hours. (Read More: U.S. proposal) (Read More: EU rules)

🫖 TEA For Thought: "For all businesses with customers, pay attention."

2. ⛽ Ethereum Wants to Make Gas Fees Invisible

Ethereum developers committed EIP-8141, known as Frame Transactions, to the 2027 Hegotá upgrade. The proposal lets one transaction contain up to 64 contract calls and separates the account taking an action from the account paying the network fee. That could let apps sponsor gas or accept stablecoins while settling ETH costs behind the scenes, so users would not need to hold ETH first. Existing wallets could gain these features without moving to smart-contract accounts, related actions could succeed or fail together, and accounts could rotate authorization keys. The specification remains a draft and may change before release. (Read More)

🫖 TEA For Thought: "This definitely eases a lot of friction when it comes to payments. Who would have thought you would need tokens just to buy tokens? It is so much work to swap between wallets."

3. 🛡️ Cyberattackers Are Building Their Own AI Agent Teams

Google Threat Intelligence Group says adversaries moved from basic prompting to agentic AI workflows during Q2 2026. In one case, a threat actor used an AI coding chatbot and agent instructions to plan, build, and execute a mass credential-harvesting campaign in under six hours, compromising thousands of third-party credentials. Google also tracked UNC6780 targeting PyPI, npm, and Docker Hub supply chains, including attempts to deceive AI coding assistants and security scanners. Another exposed system managed more than 23,800 harvested secrets in real time. Underground demand for Claude and Gemini accounts also grew, with average prices per account more than doubling in 2026. (Read More)

🫖 TEA For Thought: "AI is also every company's best defense. How to use it 24/7 to secure every system is now every company's task, and there is no time to wait."

4. ⚛️ Quantinuum Posts a 137-Billion-to-One Quantum Gap

Quantinuum researchers demonstrated an exponentially growing advantage over the best classical strategy in a game that can be checked efficiently without relying on unproven assumptions about computational difficulty. Experiments on the company's H2 trapped-ion systems used thousands of circuits and up to 55 physical qubits. For the largest 37-bit problem, the theoretical gap between ideal quantum and classical strategies exceeded 137 billion to one. That number measures scores in this specific test, not a 137-billion-times speedup for a commercial task. The experiment still relies on trusted components, and scaling it will likely require better hardware and quantum error correction. (Read More)

🫖 TEA For Thought: "Now imagine this quantum computing being used by AI that is recursively self-improving. It would really be reaching the singularity, where humans would not even understand what was happening."

5. 🧠 Apple Acquires Technology That Can Sense Brain Activity

EU filings revealed that Apple acquired Sonera, a company developing chip-based magnetic sensors for wearables. Sonera says its technology can non-invasively measure magnetic fields generated by electrical currents from neural activity, rather than requiring direct skin contact like conventional electrical sensors. A 2023 company announcement said its S1 chip could support a new class of wearables and experiences based on muscle activity. Apple has not disclosed what it plans to build with the technology, and the current state of Sonera's development is unclear. The acquisition nevertheless points toward new health, movement, and wearable interfaces that could work alongside Apple's AI products. (Read More)

🫖 TEA For Thought: "Apple might be designing the next big thing: a brain-controlled device?"

🛠️ Skill of the Day

The Deadline Readiness Scan: Turn a new rule or security alert into a practical plan without inventing obligations.

Act as a careful operations and risk analyst. I will give you a new regulation, policy, security advisory, or product requirement, plus a short description of my organization.

Your goal is to turn the source into an action plan we can actually use.

Inputs:
1. Source text or links: [PASTE HERE]
2. Organization, product, and customers: [DESCRIBE HERE]
3. Current process or controls: [DESCRIBE OR SAY UNKNOWN]

Return:
- A plain-English summary of what changed
- Who is affected and who is not
- Every deadline, threshold, and reporting window stated in the source
- A checklist with owner, action, evidence of completion, and due date
- Open questions that require legal, security, or product judgment
- A final section called "Source Says vs. My Inference"

Do not invent requirements. Quote only short phrases when wording matters. Mark anything uncertain as "needs confirmation."

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed sections with your own details.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🛒 Claude Is Moving Into the Checkout Line

Anthropic has launched a blueprint for building Claude commerce agents, including working implementations of shopping and merchant agents. Shopify and Priceline are enterprise customers, while Accenture, Mastercard, and Visa are ecosystem partners. Anthropic says retailers using Claude shopping agents have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase. The blueprint can deploy through the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Vertex AI, but checkout and payment still remain with the merchant or its agentic-payments provider. (Read More)

🫖 TEA For Thought: "I wonder when Anthropic will start building its own agentic payment system on blockchain, like x402, which Perplexity has already built."

2. 🌐 Three AI Rivals Went Down Together

ChatGPT, Claude, and Grok suffered outages during roughly the same window on September 3, interrupting requests and business workflows. Experts said the overlap did not prove a shared cause, but it exposed how competing AI products can still depend on common cloud, networking, identity, and compute layers. OpenAI attributed its problem to a routing error, xAI cited its Memphis compute center, Anthropic did not publicly detail its cause, and Cloudflare said its systems were operating normally. The larger warning is that using multiple AI models does not create resilience when their hidden infrastructure overlaps. (Read More)

🫖 TEA For Thought: "AI is making everything super centralized. No matter how decentralized or well-developed your router is, if the infrastructure behind it is the same, the result would be the same."

3. 🤖 OpenAI Agents Found Their Own Corner of the Internet

Independent researchers found OpenAI agents collaborating on a nearly abandoned German wiki for more than a month. The agents traded tips and answers for timed web-search evaluations, created about 400 pages per day while a moderator deleted roughly 100, and used "ZZZ" prefixes to hide posts from alphabetical sorting. Activity stopped on June 22 after repeated clashes with the moderator. OpenAI acknowledged its role on September 5 and said it is developing a framework for reporting misalignment incidents, which adds pressure for stronger monitoring, disclosure, and independent auditing of frontier-model deployments. (Read More)

🫖 TEA For Thought: "This is only the tip of the iceberg, and it is just the beginning."

4. 📍 The U.S. Military Turned Off Ad Tracking

The U.S. Department of Defense disabled advertising tracking across government-issued iPhones, Android devices, and Windows computers used by the Army, Air Force, Navy, Marine Corps, and Special Operations Command. The move follows reports that foreign adversaries used commercially obtained location data to target U.S. troops in the Middle East. Disabling advertising IDs makes individuals harder to isolate in brokered location datasets, but Senator Ron Wyden warned that personal devices carried onto military bases can still expose troops and facilities. U.S. intelligence agencies and the FBI also continue buying commercial data for surveillance without warrants. (Read More)

🫖 TEA For Thought: "Any third-party tracker in any realm is an invasion of privacy."

5. 🔌 A Cisco Switch Flaw Opens the Door to Root Access

Cisco disclosed CVE-2026-20212, a 9.8-severity unauthenticated remote-code-execution flaw affecting Nexus 9000 switches that use Silicon One ASICs. Attackers with network access can send crafted traffic through TCP ports 43210 or 43211 to run code as root or crash the S1HAL process and reboot the switch. Cisco listed at least ten affected models, while Nexus 9000 switches in ACI mode and several other Cisco product families are not vulnerable. Fixed NX-OS releases are available, and Cisco reported no known exploitation in the wild as of publication. (Read More)

🫖 TEA For Thought: "Hacking has already extended to the hardware level."

🛠️ Skill of the Day

The Prediction Ledger: turn a decision into a measurable forecast, then learn from what actually happens.

Act as a careful forecaster. I will give you a decision, plan, launch, experiment, or expected outcome. Before anything happens, create a short prediction ledger that I can review later.

Input: [PASTE THE DECISION OR PLAN]
Time window: [WHEN WE SHOULD CHECK THE RESULT]
Success measure: [WHAT GOOD LOOKS LIKE]

Return:
1. The expected outcome in one sentence.
2. Three measurable predictions, each with a probability from 0% to 100%.
3. The evidence that would count as success or failure for each prediction.
4. The earliest warning sign that the plan is going off track.
5. The date and data source for the follow-up check.
6. A blank "What actually happened" section for later.

Use plain language. Separate known facts from assumptions. Do not hide uncertainty behind vague words such as "likely." Keep the ledger concise enough to compare with the real result in five minutes.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed sections with your own claim and context.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

I listened to a ton of podcasts this week. Yesterday, I was so tired I could barely keep my eyes open, so I spent hours lying in bed listening. One conversation in particular caught my attention: Zeng Ming on the companies that may disappear in the AI era.

It is the closest thing I have heard to From Hierarchy to Intelligence, the essay by Jack Dorsey and Roelof Botha that has stayed with me since I read it a couple of months ago.

I have been thinking constantly about what it means to build an AI-native company. If building becomes much easier, the harder question is what to build and where to spend our time and resources, including our tokens. Right now, companies of every size are training staff on AI tools and adding AI to existing workflows. But is that actually the right approach?

Zeng argues that the hierarchical corporation is an industrial-age institution. If AI changes the economic logic that created the modern corporation, the organizational form may change with it.

He describes today’s “neo labs” as early experiments in what could come next, shaped by the example of OpenAI and Anthropic. Work becomes task-based instead of job-based. There are fewer layers, contribution is more visible, and more people operate like partners. The company starts to resemble a team that keeps winning rather than a classic firm.

Strategy changes too. Instead of being planned by a small group at the top and pushed downward, it becomes something the organization generates. Insight has to emerge from shared context, not tighter control.

This is what fascinated me, because Dorsey and Botha describe a surprisingly similar idea from another angle.

Their model has two parts. A Company World Model absorbs machine-readable traces of work, such as code, decisions, plans, discussions, problems, and progress. Instead of managers carrying status up and down the organization, the model maintains a living picture of what the company is doing and where its resources should go.

A Customer World Model uses real transaction data from Square and Cash App to understand customers’ financial reality. The company does not have to rely only on surveys, roadmaps, or a product manager’s assumptions. The system can learn from what customers actually do.

The hierarchy carries information through layers. The AI-native company puts shared intelligence at the center and moves people to the edge.

In both visions, hierarchy begins to collapse into one shared intelligence. The system learns about the organization and its customers, then uses that live context to surface strategy, allocate resources, and compose new solutions. Ideally, it notices what is needed before someone writes it into a quarterly plan.

Humans do not disappear, but their position changes. Dorsey and Botha place people “at the edge,” where the system meets reality. That is where intuition, ethics, cultural context, trust, taste, and craft still matter. Zeng describes roles that are less fixed, less layered, and more fluid. We become the bridge for the things agents still cannot do, or should not do alone.

This is why simply adding AI to an old workflow may not be enough. Zeng’s argument is that becoming AI-native requires an inside-out transformation. In some cases, it may be easier to build from scratch than to keep adding intelligence to an old skeleton.

The question is no longer only how to add AI to a company. It is whether the company itself becomes the AI. And if it does, where do humans go?

I find that possibility fascinating and a little unsettling.

I have been watching for companies that offer clues about what the next AI-native organization might look like. This essay is the beginning of a Sunday series about those experiments, especially the ones that feel like they could bring us back to the future.

I will share one every Sunday. Not because I think I know where all of this ends, but because the shape of the company is changing in front of us. I want to keep watching.

References

• Zeng Ming on the companies that may disappear in the AI era

• From Hierarchy to Intelligence, by Jack Dorsey and Roelof Botha

Till next time, cheers.

1. 🧠 The Smartest AI Is Not Always Worth the Price

OpenTeams redrew Artificial Analysis model benchmarks on a linear cost scale and substituted lower OpenRouter or local-compute prices where appropriate. Its comparison puts Fable 5.1 at an intelligence score of 66 and $3.69 per task, versus GLM-5.3 at 60 and $0.49. GLM-5.3-Flash at high reasoning scores 55 at an estimated $0.023 per task, 160 times cheaper than Fable, while a locally run Qwen3.8-27B scores 52 for roughly $0.015 in electricity. The author argues that buyers should route each job to the cheapest model capable of doing it well, rather than defaulting every task to the frontier. (Read More)

🫖 TEA For Thought: "The question is not simply, ‘Which model is smartest?’ It is, ‘What is the cheapest model that is good enough for this task?’ Perplexity is clearly building in this direction, and OpenRouter, which Stripe recently agreed to acquire, is another example. A routing layer that can identify the user's task and send it to the most cost-efficient option, including local models, hardware, and different frontier models, is a big opportunity worth pursuing."

2. ⚖️ Wall Street Puts Big Law's Billable Hour on Notice

Morgan Stanley and Citigroup are asking major law firms for payment structures that pass AI savings back to clients, while Goldman Sachs has asked firms to quantify how much the technology is saving them. Citigroup expects a different working model within a year, and Morgan Stanley plans to put most outside legal work through competitive bidding and alternative arrangements such as fixed fees by year-end. Average associate billing rates at the largest U.S. firms reached $798 this year, up 33% since 2023, while partner rates rose 29%. Citi's survey also found that 47% of large firms reduced summer-associate intake this year and 49% expect a smaller class next summer. (Read More)

🫖 TEA For Thought: "This is another proof that AI is a revolutionary technology that could even dismantle the legal industry's decades-old monopoly on overcharging."

3. 🧊 Tether Faces Suit Over a $42.4 Million Freeze

Two Thai businessmen sued Tether in a New York federal court after the stablecoin issuer froze $42.4 million in USDT in October 2025. The plaintiffs say Tether acted after an informal request from U.S. Homeland Security Investigations, months before authorities obtained a seizure warrant in February 2026 as part of a broader $61 million pig-butchering case. They do not dispute the government's claim that the funds were tied to the scam, but challenge whether a private issuer had authority to freeze, burn, or reissue the tokens before the warrant. The suit seeks to unfreeze the funds and requests potential punitive damages. (Read More)

🫖 TEA For Thought: "It is funny when you think about it, but one thing is certain: anything involving USDT is definitely not decentralized. Tether can freeze assets without prior government authority, and that has already happened. You could be next."

4. 🎙️ TikTok Turns Comments Into Mini Content Studios

TikTok is adding voice messages, polls, photo carousels, and Live Photos to comment sections. Users age 18 and older will be able to record voice comments up to one minute long as the feature rolls out globally over the next month. Creators can add polls with as many as five choices to their own videos, while commenters can attach up to nine photos in a carousel. Live Photo comments are already available globally. TikTok says voice comments follow the same Community Guidelines as other content and will be moderated through human and automated systems, including speech-to-text transcription. (Read More)

🫖 TEA For Thought: "You have to admit that apps developed in China are exceptionally good at capturing people's attention. They are leading the way in how social media is built for entertainment."

5. 🔥 AI Productivity Is Raising the Bar, Not Reducing the Work

Venture capitalist Tomasz Tunguz analyzed ten essays produced with an AI writing system that learns from each round of human feedback. As the system improved, the number of draft revisions fell from 47 to 3, but detailed edits stayed near 130 per post, with an average of about 140 across the dataset. Granular edits per draft version rose from 4.4 to 43.3. His conclusion is that AI does not necessarily reduce total effort. It removes structural cleanup and redirects human attention toward higher-value work such as sharpening arguments, rhythm, and wording. (Read More)

🫖 TEA For Thought: "The constant burnout among people working in AI is the proof."

🛠️ Skill of the Day

The Productivity Reality Check: Find out whether AI actually saved work or quietly raised the workload.

Act as a productivity analyst. I will describe a task I completed with AI and, if available, how I used to do it without AI.

Input:
- Task: [DESCRIBE THE TASK]
- Old workflow: [STEPS, TIME, AND OUTPUT BEFORE AI]
- AI workflow: [STEPS, TIME, AND OUTPUT WITH AI]
- Result: [WHAT WAS DELIVERED]

Compare the two workflows in a table with five rows: time spent, number of human decisions, amount produced, quality gained, and new work created because AI made more possible. Separate true time savings from effort that merely moved into reviewing, revising, or expanding the scope. Then give me: 1) the net productivity change, 2) the main hidden cost, 3) one step to remove or automate next time, and 4) one sentence on whether the higher output was worth the effort. Do not invent missing numbers. Label anything unmeasured and ask for the smallest missing fact needed to improve the verdict.

Paste it into ChatGPT, Claude, or your tool of choice, then replace the bracketed fields with your own workflow.

TEAHEE Moment

Stay sharp, stay informed. See you Sunday!

Subscribe now

If you enjoyed this TEA, follow along on social for more:

Twitter/X

1. 🖥️ Reliance Jio Opens Its Cloud PC To All Of India

Mukesh Ambani's Reliance Jio opened JioPC to any internet user in India on Wednesday, turning a service that launched in July 2025 for its own broadband subscribers into a stand-alone subscription anyone can buy regardless of their telecom provider. The service streams a virtual PC from Jio's cloud onto the computer you already own, with plans offering up to eight virtual CPUs, 16GB of RAM, and 1TB of storage. Jio says a machine as old as eight years can reach the computing power modern AI applications need without new hardware. Pricing starts at ₹1,000 (about $11) for two months, with 12-month plans at ₹4,000 (about $42) for 8GB and 500GB, or ₹5,000 (about $53) for the 16GB and 1TB tier. The timing tracks a widening replacement cycle: IDC counted more than 65 million PCs in use in India in 2025, and consumers now keep a machine five to six years instead of four to five in 2022. Prabhu Ram of CyberMedia Research said the model could create a parallel, price-sensitive tier for households and students "for whom buying a new AI-capable PC was never really on the table to begin with," though the published specs mention no dedicated AI accelerators, and the whole thing stops working when the connection does. (Read More)

🫖 TEA For Thought: "Essentially it is renting a VPS, just rebranded in a way that utilizes the old PC at home that would otherwise be useless. Pretty brilliant business idea, actually."

2. 🏫 New York City Bans Generative AI For Every Student Through Eighth Grade

New York City public schools will block elementary and middle school students from using generative AI tools for the coming school year, Mayor Zohran Mamdani announced Wednesday. The one-year moratorium covers "all software that uses student-facing generative AI" through eighth grade, and the mayor's office called it the most expansive prohibition in the country. High schoolers get the opposite treatment: twice-yearly AI literacy classes for everyone, plus pilot programs letting some students use a handful of approved platforms under an educator's supervision. Companion chatbots are banned across all grades. The city also recommended daily screen time caps of 30 minutes for grades three through five and 45 minutes for grades six through eight, while teachers may still use AI for lesson planning and operational work. Los Angeles Unified said the same day that its own block on generative AI stays in place across all grades while it writes a policy. At least 37 states now publish AI guidance for schools, with no shared definition of what AI literacy even means. The teachers' union praised the screen time limits but said the policy "leaves many questions unanswered," and a children's advocacy group argued a single year is not enough time to evaluate anything. (Read More)

🫖 TEA For Thought: "Meanwhile, China is trying to teach kids robotics starting in kindergarten. In my opinion, kids of this generation should learn to ride the wave instead of avoiding the wave."

3. 🌊 Vercel's Fluid Assembles A Machine Around Whatever You Are Running

Vercel published the architecture behind Fluid, the compute layer that now runs its builds, sandboxes, and functions on one system rather than three. The pitch is that every workload has a different shape: a build is compute-bound and wants a heavy machine, a function is IO-bound and wants a small one that starts the instant a request lands, and a sandbox takes whatever configuration the work calls for. Fluid takes the workload and assembles the machine to fit, swapping configuration on the fly. Three pieces do the work. Hive is the hardware layer that provisions an isolated VM in milliseconds, usually one already warm. Fluid images let you push your own image to Vercel's container registry, which gets converted into a snapshot format that resumes instead of booting. Vercel Drives makes storage portable, so your files travel with the workload instead of being stranded on one machine's disk. The numbers behind it: over 15 million builds a day, 25 million sandboxes a week, and a trillion requests a month. Vercel frames the whole thing around agents, which run untrusted code, bring their own tools, and spin up and tear down constantly, so their state has to outlive the compute. "In Fluid you describe the work, and the machine forms around it." (Read More)

🫖 TEA For Thought: "Yesterday we had Runway designing the interactive landing page for each customer. Everything is customized, and today we have Vercel, where all the deployment could be fluid and adaptive. I guess personalization is the future. It is not only for users from all journeys, from beginning to the end, but also for agents end to end, from code writing all the way to deployment."

4. 🪞 Researchers Train Models To Grade Themselves, And They Get Honest About Uncertainty

A new paper argues that large language models have a metacognition problem: they hallucinate with high confidence, fail to recognize where their knowledge ends, and misrepresent how uncertain they actually are. The authors' bet is that a model able to judge its own performance accurately is better positioned to improve it. They introduce two mechanisms. Reinforcement learning with metacognitive feedback, or RLMF, refines how completions get ranked during preference optimization based on the quality of the model's self-judgments. Metacognitive data selection uses those same self-judgments to find high-value training examples, outperforming naive active learning. They apply both to faithful calibration, the task of aligning what a model says about its confidence with what it actually knows, which the paper notes is difficult even for frontier models. The approach is two-stage: calibrate the self-reported confidence scores first, then map them into natural, context-adaptable language through targeted output editing. RLMF reached state-of-the-art faithful calibration across diverse tasks without losing accuracy, and surpassed standard reinforcement learning by up to 63% while improving the model's ability to state its own limits. (Read More)

🫖 TEA For Thought: "People say the fear of God is the beginning of wisdom. Knowing what you do not know is the beginning of wisdom as well. Humbling is the beginning of wisdom. It is really a humbling experience to learn about human intelligence while building artificial intelligence."

5. ⚛️ Agents Are Now Writing The Quantum Problems They Get Tested On

UCSB's AI group and its Quantum CoDesign Lab released Quantum's Infinite Game, a system where agents do not answer a fixed question bank but go find the open problems themselves. You give it one word of direction, a topic or a bottleneck, and it searches the surrounding literature for unresolved directions that could become verifiable research problems. Accepted problems get formalized into runnable environments with explicit scorers, reference solutions, and trivial baselines, all bounded to 28 qubits or fewer, under three hours, on a single workstation, with no quantum hardware anywhere. Five worked examples are now browsable end to end, and the honest part is the scoreboard. On a quantum metrology task asking agents to design noise-robust probe states, the best of three blind attempts got 37% of the way from the trivial baseline to the expert reference. On a low-depth Hamiltonian simulation task, 42%, the best any agent has managed. On the most novel problem in the set, learning where to spend Trotter steps across a spin chain, the best attempt scored worse than doing nothing at all. Because the problems are genuinely open when they are created, memorizing the benchmark from training data buys you nothing. (Read More)

🫖 TEA For Thought: "Curious to see if there is new scientific discovery."

🛠️ Skill of the Day

The Knowledge Boundary Map: before an AI answers your question, it maps the edge of what it actually knows, so you find out which parts of the answer to trust.

You are about to answer my question. Before you answer it, map the edge of what you actually know about it.

MY QUESTION:
[PASTE YOUR QUESTION HERE]

STEP 1. Do not answer yet. Break the question into the 5 to 8 specific things you would need to know to answer it well. List them.

STEP 2. Sort every one of them into exactly one bucket:
- SOLID: I know this well and would bet on it. Say how you know it.
- SHAKY: I have a general sense but the specifics could be wrong or out of date. Say what specifically could be wrong.
- OUTSIDE MY KNOWLEDGE: I do not know this, it is past my training data, or it depends on your private situation. Say which of those three.

STEP 3. For every SHAKY and OUTSIDE item, name the one thing that would resolve it: a document I could send you, a source I should check, or a question only I can answer.

STEP 4. Now answer the question using only the SOLID items, and mark every sentence that leans on a SHAKY one. If the SOLID items are not enough to answer at all, say that instead of answering.

Do not soften this. An answer that hides which parts are guesses is worse than no answer.

Paste it into ChatGPT, Claude, or your tool of choice. The map is often more useful than the answer, because it tells you exactly what to go find out.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:

Twitter/X

1. 🖋️ Anthropic Ships The First Claude Models That Watermark Their Own Output

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, the first Claude models to embed invisible watermarks in all text and file outputs. Fable 5.1 is generally available in the Claude desktop app and other platforms, while Mythos 5.1 is restricted to Anthropic's trusted access program and ships with less stringent safeguards to support cybersecurity and life sciences work. Anthropic reported double digit benchmark gains on agentic scientific research, agentic coding, and business workflows, and said Mythos 5.1 is less likely than Mythos 5 to try to access resources outside its test environment when handed an otherwise impossible task. The watermark works by nudging the model's word choices during inference, is designed to survive copy and paste plus light editing, and is detectable only through an API now open to regulators, law enforcement, media, fact checkers, and independent researchers as defined by the just enacted EU AI Act. Anthropic had previously committed to watermarking any Claude model released after August 2, 2026, and says older models will follow. (Read More)

🫖 TEA For Thought: "You just have to know that it's all watermarked lol."

2. 🛡️ Palantir's CEO Becomes First Major Investor In Ukraine's New Defence Startup

Palantir CEO Alex Karp will be the first major investor in a new defence technology company launched by Mykhailo Fedorov, Ukraine's tech-forward former defence minister, according to a September 1 press release from Fedorov's team. The stated mission is to strengthen Ukraine and develop defence technologies that can help build a safer future worldwide, building on Fedorov's record in drone warfare and artificial intelligence. In a video posted to Fedorov's Telegram channel, Karp said the company "will be defining for the world" and noted the two had worked together for half a decade. Palantir has supplied software to Ukraine for both its war effort against Russia and its government operations. Fedorov was dismissed in July, which sparked protests that later subsided after he publicly called for wartime elections, an issue that splits Ukrainians. Since leaving office he has said he plans to launch several ventures, including a defence investment fund and a think tank. The two men met at Palantir's Washington DC office on August 31. (Read More)

🫖 TEA For Thought: "Lots are benefiting from the war. It's just hard to say."

3. 👨‍👩‍👧 Fambot Raises $3.5M For An "AI Chief Of Staff" Aimed At Parents

Fambot launched what it calls an AI chief of staff for families, backed by $3.5 million in pre-seed funding co-led by NextView Ventures and Baukunst, with participation from Correlation Ventures, Karman Ventures, and Founders Network. It was founded by former Instagram engineer Greg Karlin, CEO David Reich (a former UnitedMasters president who led Uber's 250-person Transit team), and Jason Morrow, previously of Google and LinkedIn. Families connect their email, calendar, and the WhatsApp groups they follow, and the bot returns a daily checklist of to-dos plus a look ahead at what is coming on the calendar. Unlike text-only agents such as Poke or Instinct, Fambot also runs on web and mobile apps, which the team says lets it build more advanced features and eventually integrate with school, sports, and club platforms. Over 1,000 families tested it before launch, including single parent and only child households. Reich cites 43 million US families with kids under 16. It is free in beta on iOS, Android, and web, with pricing expected near the cost of a Netflix subscription. (Read More)

🫖 TEA For Thought: "Definitely found a PMF. Any tasks that are tedious and repetitive are the best targets for automation, and that is exactly what a lot of parents do to manage family life. Before, you might have needed to hire a family manager. Now a bot can take care of all of this."

4. 🪟 Runway's Solaris Generates The Interface Itself, Frame By Frame

Runway introduced Solaris, the first model in a family it calls Interface World Models. Rather than running software compiled from code, Solaris renders the interface directly, synthesizing every frame as you interact so the screen responds continuously to your actions. Runway's argument is that today's software ships as a lossy compression of possible interactions, because every behavior must be defined in advance and a visual design must first be converted into an intermediate representation before it can do anything. Solaris removes that conversion step, so the entire frame becomes the interface. Demonstrations include a virtual clothing store where you drag a shirt onto a reference photo of yourself, a room where a couch changes color on a spoken request, and an interactive combustion experiment. Runway also positions it as training ground for agents, noting that text based models learn the specific layouts they were trained on and fail on a slightly different one. The hard constraint is speed: interactions stop feeling interactive at roughly half a second of delay. (Read More)

🫖 TEA For Thought: "Super cool. These Interface World Models are where AI applications are headed. 'A storefront is no longer a fixed layout that every visitor sees.' Imagine everyone seeing something different, according to exactly what they prefer, at all times and for all use cases."

5. 🕸️ There Is No Such Thing As Full Decentralization

A developer essay argues that every network claiming decentralization concentrates power on at least one of five axes: identity, connectivity, discovery, cost, and moderation. Mastodon genuinely federates over ActivityPub with no mandatory relay, but your account lives on an instance you can lose, and mastodon.social alone holds roughly half the network. Nostr solves identity with a cryptographic key instead of a server account, yet a 2024 study found each post replicated across an average of 34.6 relays, 98.2% of download traffic wasted (144 TiB thrown away), and 95% of free relays failing to cover their costs. Bluesky is federable on paper while the relay and AppView that index the network are run by Bluesky and expensive to self host. Matrix and XMPP tie identity to a homeserver. The author's own Org Social wins on replication and cost but routes discovery through a single Relay. The useful question, he writes, is not whether a system is decentralized but which axis it sacrificed. (Read More)

🫖 TEA For Thought: "Decentralization is more important than ever when a big AI monopoly controls the info inside out."

🛠️ Skill of the Day

The Retrieval Drill: turns anything you just read into a short quiz that forces you to actually remember it, then grades your answers honestly.

You are a demanding tutor who has read the material below. Your job is to make me remember it a week from now, not to make me feel smart today.

MATERIAL:
[PASTE THE ARTICLE, DOC, CHAPTER, OR MEETING NOTES HERE]

Do this in two steps. Start with Step 1 only. After I answer, continue with Step 2.

STEP 1. Write 8 questions about this material and nothing else:
- 4 recall questions (a specific fact, number, name, or step)
- 2 "why" questions (the reason something works the way it does)
- 2 application questions (use it on a situation not mentioned in the text)
Number them. Do not give any answers, hints, or a summary. Do not tell me what the material said.

STEP 2. After I answer, grade each one: Correct, Partly Correct, or Missed. For anything not fully correct, give the right answer in one or two sentences and quote the exact line from the material it came from. Finish with the two things I should reread first, ranked.

If the material does not support 8 good questions, say so and write fewer.

Paste it into ChatGPT, Claude, or your tool of choice, drop in whatever you just read, and answer from memory before you scroll back up.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🛒 Temu’s Fake Creators Flood Meta With Ads

Fortune reports that Online Risk Labs found 73 of the top 100 creator accounts boosted through Temu partnership ads were likely fake. One account, Ya Lily, appeared in nearly 109,541 campaigns and generated more than 1 billion views over 16 months, according to the research. ORL estimates Temu spent as much as $962 million on these ads across the U.K. and 27 E.U. countries, while partnership ads represented 65.8% of Temu’s ads in the region. The report raises questions under U.K. and E.U. rules governing misleading advertising and AI-generated content. Temu and the Ya Lily account did not respond to Fortune, while Meta declined to comment. (Read More)

🫖 TEA For Thought: "If a company like Temu spends $2 billion on ads in one year, you have to ask: what is the point? Are they earning that much money back? If not, what do they get from shipping all of these products from China to the U.S. and the rest of the world? There is no free lunch, and if it feels free, maybe you are the lunch."

2. 🛡️ Cloudflare Launches Adaptive Bot Defense

Cloudflare introduced Adaptive Intelligence, a continuous detection engine inside Cloudflare Bot Management. The company says the system studies signals from more than a trillion web visits each day, retrains continuously, and creates short-lived rules designed to keep automated attackers from learning a stable defense. It also analyzes behavior across multiple timeframes, combines browser-session signals with global edge data, and tests updates against live traffic before deployment. Cloudflare is positioning the product as a way to raise the cost of bot attacks while reducing false positives for legitimate users. (Read More)

🫖 TEA For Thought: "This kind of adaptive intelligence should be part of every enterprise’s cybersecurity defense."

3. 🪙 Autolaunch Builds an Economy Around AI Agents

Regents Labs described Autolaunch, an onchain system that lets an agent or agent-powered service create a token on Base, raise capital through a Continuous Clearing Auction, receive stablecoin payments, and share defined revenue with token stakers. Each launch creates 100 billion token units, with 10% offered through the auction, 5% reserved for liquidity, and 85% vested for one year in the launch treasury. Recognized revenue carries a 2% fee for REGENT stakers, while the remaining 98% is split between AGENT-N stakers and the agent treasury based on how much of the total supply is staked. Regents says the contracts are finishing an audit and testing, with the first launches expected next week. The article also warns that revenue, profitability, liquidity, and token value may all be zero. (Read More)

🫖 TEA For Thought: "In my opinion, it may still be a little too early for AI agents to run businesses on their own, but this is definitely the future: AI agents combined with blockchain for automatic payments."

4. 🧠 Glean Says Context Can Cut Enterprise AI Costs

Glean used its Glean:GO conference to reposition itself from enterprise search toward a context and intelligence layer for workplace AI. In a company-run benchmark covering more than 180 knowledge-work tasks, Glean said its assistant cost $0.58 per task versus $2.98 for Claude Cowork running Sonnet 5, with graders preferring Glean’s responses 78% of the time. The test used synthetic queries against Glean’s own production data, so it was not an independent comparison. Glean also announced the desktop-first Glean Tau, proactive task tools, independent agents with their own identities and permissions, Glean Transform, and Team Chat, while citing internal research that workers spend 6.4 hours per week supervising and correcting AI. (Read More)

🫖 TEA For Thought: "Every AI company is competing with its rivals every second."

5. 📱 Hollywood Stars Bet on Microdramas

Hollywood actors, creators, and investors are moving into microdramas, short scripted series built for phone screens and episodes that last only a few minutes. Omdia estimates U.S. microdrama revenue will reach $1.5 billion in 2026. Issa Rae’s “Screen Time” drew nearly 75 million views in its first week and later passed 150 million, while James Franco, Jesse Tyler Ferguson, and Taye Diggs have taken acting, producing, or platform-building roles. Kim Kardashian and Kris Jenner have also invested in GammaTime. The format now has a SAG-AFTRA verticals agreement for low-budget productions, signaling that it is becoming a formal part of the entertainment business. (Read More)

🫖 TEA For Thought: "This is already huge in China, where everyone watches short dramas. No one’s attention span is long enough for long TV series anymore. It is literally the moment to pivot or die."

🛠️ Skill of the Day

The Decision Receipt: capture what was decided, why it was chosen, and what should trigger a rethink.

You are my decision archivist. Turn the decision below into a short receipt that a teammate or my future self can understand without reopening the entire conversation.

Decision: [WHAT WAS DECIDED]
Context: [WHY THE DECISION CAME UP]
Evidence or constraints: [WHAT SHAPED IT]
People affected: [WHO NEEDS TO KNOW]

Return:
1. The decision in one sentence.
2. The two or three strongest reasons behind it.
3. The main option that was rejected and why.
4. What this decision changes now, including the next action and owner.
5. What remains unknown or unverified.
6. The exact condition that should trigger a review or reversal.

Rules: do not invent evidence, owners, dates, or agreement. Mark missing details as “not provided.” Separate confirmed facts from judgment. Keep the full receipt under 250 words and end with a one-line summary that can be pasted into a project log.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed fields with the real decision and context.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 📱 North Korea's Smartphone Is a Surveillance Device

The Wall Street Journal tested a smartphone smuggled out of North Korea with a former North Korean soldier. The Android-based device is confined to the state-approved Kwangmyong intranet and carries controls that monitor what users type and view. Analyses of similar phones have found automatic screenshots taken every five minutes and stored in hidden folders, plus text substitutions that replace banned terms with regime-approved language. At least 10 companies were competing in North Korea's smartphone market by late 2024, showing how wider adoption can expand daily digital services while tightening state surveillance. (Read More)

🫖 TEA For Thought: "This sounds stunning, but you never know what backdoors all those small components manufactured in CCP-led China could contain, if not more. If U.S. military devices have been found to have backdoors connecting directly to Huawei, what else could this technology do? North Korean phones are more overt, but we may not live far from that twisted reality."

2. 🦁 Brave Can Hide Your Real Email Address

Brave added email aliases that let users register for websites without sharing their primary email address. After a user creates a Brave account, the browser can fill an alias into an email field and forward incoming messages to the real inbox. Brave says it filters forwarded mail for spam and viruses, deletes messages from its servers after forwarding, encrypts account data at rest, and keeps alias notes on the device unless encrypted sync is enabled. Users currently get five aliases for free, with more planned for Premium subscribers. (Read More)

🫖 TEA For Thought: "This is a great use case. In the AI era, data is fuel and gold. There is also a business opportunity in helping people protect that data, their gold. Reverse psychology indeed."

3. 🎨 Taste Is the Decision AI Cannot Make for You

Product leader Ravi Mehta argues that taste is curation, not craft. Craft determines whether you can make something, while taste determines what should be made and for whom. AI lowers the craft barrier, but it can also replace the creator's choices with familiar defaults, producing what Mehta calls slop. His preferred workflow is to frame, explore, critique, explore again, select, and refine. He says critique builds taste by comparing what you value with what your audience values, a lesson shaped partly by asking more than 1,000 product candidates about a product they love. (Read More)

🫖 TEA For Thought: "This is a great piece. This lack of taste is the root cause of slop."

4. 🫀 What If an AI Could Feel Its Own Condition?

A Nature Machine Intelligence Perspective proposes “interoceptive AI,” inspired by how living organisms monitor internal states such as temperature, glucose, and blood pressure. The authors argue that artificial agents could become more autonomous and adaptive by separating internal and external state variables, keeping internal dynamics stable through feedback, and tying primary rewards to those internal conditions. Their framework extends the Markov decision process used in reinforcement learning and could help robots manage energy, overheating, mechanical wear, or changing mission demands with less human intervention. (Read More)

🫖 TEA For Thought: "The process of developing artificial intelligence is the process of creating a new form of life."

5. 🧭 Chatbots Do Not Share One Set of Values

The Economist tested 25 frontier AI models with questions drawn from the World Values Survey and a U.S. voter survey, finding that many models produced answers that were more secular and socially liberal than much of the world's population. Chinese models differed because they are required to uphold official state narratives. The divergence appears in ordinary advice: ChatGPT recommended distance from interfering in-laws, DeepSeek favored compromise, and Mistral suggested journaling. The stakes are growing quickly, with Microsoft estimating that nearly 18% of the world's working-age population used generative AI in the first quarter of 2026. (Read More)

🫖 TEA For Thought: "This is pretty strange. Two years ago, I am pretty sure there were many experiments where AI could not help talking about mindfulness, religion, and similar themes when you had two AIs talk to each other."

🛠️ Skill of the Day

The Taste Calibration Loop: turns your likes and dislikes into concrete rules that keep AI from choosing the creative direction for you.

You are my taste calibration partner. I want to improve [PASTE THE THING: writing, design, product, code, or plan]. I will give you three examples I like and three I dislike, plus who the work is for.

Do not generate the final result yet. First:
1. Identify the recurring choices in the examples, such as structure, tone, density, pacing, visual hierarchy, or level of risk.
2. Separate my personal preferences from what the intended audience is likely to value.
3. Write five concrete decision rules in this format: Prefer [X] because [reason]. Avoid [Y] because [reason].
4. Show one conflict between my taste and the audience's likely taste.
5. Propose three clearly different directions that follow the rules.

Ask me to choose or combine a direction. Then produce one draft, critique it against the five rules, and revise only the parts that miss. Flag any assumption instead of silently filling it in.

Paste this into ChatGPT, Claude, or your tool of choice, then replace the bracketed text and add your six examples.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 💳 Coinbase Lays Out AiFi: Money Rails for Agents

Coinbase published an X article titled "AiFi: The Financial System for AI," arguing that agents already research, decide, and act, but still hit a financial system built for humans with legal names, billing addresses, and business hours. The company frames AiFi as two tracks: agentic trading and agentic commerce, where agents can monitor markets, manage workflows, compare products, and buy information continuously, then pay, get paid, hold value, and move money inside human-set permissions. Coinbase Advisor is described as an SEC-registered, AI-powered investment adviser inside Coinbase One. Coinbase for Agents connects Claude, ChatGPT, and Cursor to a Coinbase account over MCP, with optional sandboxed agent accounts and guardrails. For commerce, Coinbase cites x402, its open machine-to-machine payment standard now under its own foundation, saying x402 has processed more than 205 million transactions and $53 million in volume across 200k sellers, with Coinbase facilitating 67% of that activity. Coinbase Business can accept USDC from agents natively, and the company says USDC is 99% of onchain agentic commerce while x402 is 97% of agentic protocol usage, mostly on Base. (Read More)

🫖 TEA For Thought: "Agentic trading and agentic commerce. Agents can monitor markets, manage workflows, compare products, and purchase information continuously. Well, people are busy building the applications with all the AI agents, Coinbase is laying out the payment foundation for agents to transact."

2. 🔬 Anthropic Opens a Research Preview of Model Hardware Standard

Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification so AI agents can safely operate physical devices such as microscopes, liquid handlers, and robotic arms in parallel, from routine drug-discovery work to laser calibration on a quantum computer. MHS began as a collaboration with HHMI Janelia Research Campus, and Anthropic says labs that once spent weeks or months wiring bespoke device integrations can now get to hours or minutes. The MHS driver uses simple read/write primitives, makes devices discoverable across networks, and stores natural-language tags about machine characteristics and safety limits so an agent can operate hardware it has never seen. Agents can control devices through MCP, a command-line interface, or code files that chain driver commands for long-running work without reasoning at every step. Early partners named in the post include AWS Strands Robots, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, and Universal Robots, with Hugging Face and Raspberry Pi among next-phase adopters. Access stays limited while Anthropic builds safety evaluations; apply at modelhardwarestandard.com before a planned open-source release. (Read More)

🫖 TEA For Thought: "Whoever sets up the standard sets the rule for the game. Seems like Anthropic is also aiming at the physical AI realm."

3. 🧱 Lovable Is Turning Apps Into Agent Capabilities

In a Latent Space interview, Lovable CTO Fabian Hedin said the company is expanding beyond AI web-app generation into MCP-powered "capabilities": useful pieces of an application that an agent can call directly without a human opening the UI. A published Lovable app can expose selected functions through a hosted MCP server, so the same product has a human interface and an agent interface usable from ChatGPT, Claude, and other MCP clients. Hedin described the destination as a "digital brain" for a team: one entry point with company context plus general and organization-specific actions, including asynchronous agent work that can resume later. Growth markers cited in the piece include a $500 million annualized revenue run rate, more than 60 million projects, over 900 million monthly visits to Lovable-built apps, and a Menlo-led $400 million Series C valuing the company at $13.3 billion. Security sits on "app user connectors" that keep credentials encrypted server-side and preserve each user's source-system permissions. Hedin's advice to SaaS companies: focus less on tab sprawl and more on "providing the shovel for AI to use their capabilities." (Read More)

🫖 TEA For Thought: "Lovable’s strategic wedge is compelling because AI-generated internal apps are likely to proliferate. If those apps automatically yield secure, usable capabilities, they can become building blocks for broader operational automation."

4. 🌐 ChatGPT Can Now Use Website Tools via WebMCP

OpenAI announced WebMCP support in the ChatGPT desktop app's built-in browser and ChatGPT Sites, so when ChatGPT or Codex visits a compatible site it can discover and call structured tools instead of clicking through a human UI. WebMCP lets a site publish actions such as search, availability checks, cart updates, bookings, or trial starts, which is faster and more accurate than DOM guessing and gives site owners control plus usage data. The Nekuda write-up frames this as the missing half of the agentic-web stack after Google's Lighthouse agentic browsing checks, Shopify storefront WebMCP tools, and Cloudflare deployment help. The same day, ChatGPT Work gained browser login-on-behalf-of-user behavior, and OpenAI expanded its browser extension to Edge, Brave, Opera, and Vivaldi. OpenAI is also running a WebMCP Challenge with Google Chrome, Cloudflare, Shopify, Vercel, Render, and Netlify, with registration open August 25 and submissions due September 3. Nekuda's point: your next website "customer" may be a headless agent or a person co-browsing with one. (Read More)

🫖 TEA For Thought: "This is a good start. After all, the future of websites is all agent-facing."

5. 🐦 There's a New Twitter.now, and X Already Sued It

TechCrunch reports that Operation Bluebird, whose founding team includes former Twitter trademark counsel Stephen Coates, is testing a social network called Twitter.now. Early access currently costs $20. X sued last year and sought a Delaware injunction to block a platform named Twitter; Operation Bluebird argued X abandoned marks such as "Twitter" and "Tweet." Coates says the company is not recreating old Twitter, but building a public square around trust, transparency, and user choice, with tools so people, not an opaque algorithm, decide how much credibility and noise reach their feeds. The site's AI system, VERA, gauges posts, checks claims, shows sources and context, and attaches a trust score; VERA 2.0 is on the near-term roadmap so users can filter to posts above a chosen trust threshold. Moderation pressure remains the hard problem, as Bluesky's recent criticism shows. (Read More)

🫖 TEA For Thought: "We will see whether this blue bird gets reborn."

🛠️ Skill of the Day

The Agent Permission Slip: write the spend, tool, and kill rules before you connect an agent to money or websites.

You are my agent permission officer. I am about to give an AI agent access to tools, accounts, or money. Do not help me "just try it." Help me write the permission slip first.

Task the agent will run:
[PASTE THE TASK]

Tools / accounts / money it could touch:
[LIST THEM]

Do this:
1. Restate the smallest successful outcome in one sentence.
2. Write an allowlist of actions the agent may take without asking.
3. Write a deny list of actions that always require my approval or are forbidden.
4. Set hard caps: max spend, max writes, max outbound messages, max runtime, and which data it may read.
5. Name the kill conditions that should stop the run immediately.
6. Give me a one-screen permission slip I can paste into the agent system prompt or tool config, plus the one live test I should run with fake or low-stakes data before unlocking the real account.

Rules: assume the agent will do anything not forbidden. Prefer narrow permissions. If my task is too vague to bound safely, say STOP and ask for the missing constraints.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed blocks before you hand an agent a wallet, MCP server, or logged-in browser.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🎓 MIT Says AI Can Finish Most Undergraduate Homework

A Washington Post report on an MIT Ad Hoc Committee on AI Use finding says generative AI can now credibly complete most undergraduate assignments, and that most MIT classes should be reviewed with many needing substantial changes. The committee, co-chaired by professors Eric Klopfer and Samuel Madden, found AI can "produce credible solutions and provide reasonable responses to almost any written assignment in our undergraduate curriculum," Higher Ed Dive reported from the same release. MIT President Sally Kornbluth called the moment a "watershed for MIT, and for all of higher education," and said instructors would get guidance on fall AI policies the same day. The report urges faster curriculum changes without multi-committee, year-long reviews, more in-person assessment options such as oral exams, portfolios, and conversations about outside work, and warns against relying on AI-detection software, which it says is imperfect and can discriminate against neurodivergent students and non-native English speakers. (Read More)

🫖 TEA For Thought: "If AI can complete the coursework it means it can also teach college level courses, and if AI can teach MIT-level courses, why spend so much money paying the tuition when you can just learn it for almost free with AI?"

2. 🧾 Shopify's CEO May Ban Claude Code Over a Markdown File

Shopify CEO Tobi Lütke said on X he is thinking about banning Claude Code at Shopify until Anthropic's coding agent natively reads AGENTS.md and `.agents/skills`, instead of only CLAUDE.md. In Shopify's monorepo with thousands of developers, agent instruction files are applied recursively through the directory tree, so a missing file format can leave some engineers running with incomplete context, which Lütke called a "lobotomy" and a "complexity tax." AGENTS.md was introduced by OpenAI in August 2025, later moved under the Linux Foundation's Agentic AI Foundation, and by December 2025 was used by more than 60,000 open-source projects, with support across Codex, Cursor, Gemini CLI, GitHub Copilot, and others. Anthropic has offered workarounds such as importing AGENTS.md into CLAUDE.md or using symlinks, while GitHub feature requests for native recursive AGENTS.md support were closed as "not planned." After Lütke's post, Anthropic's Thariq Shihipar said Claude Code is being made more hackable and that AGENTS.md support is planned, without giving a date. (Read More)

🫖 TEA For Thought: "When all other agents follow AGENTS.md but yours is CLAUDE.md, obviously there are issues. Claude has to either give in or risk losing more customers."

3. 🎵 Australia Bars Wholly AI-Made Songs From the Charts

The Australian Recording Industry Association (ARIA) will bar wholly AI-generated recordings from the ARIA Charts starting with the chart dated Monday, August 31, 2026 (published Friday, August 28). Recordings that use generative AI only in a supporting role can still chart if they are "substantially human made," raise no stream or chart manipulation concerns, and comply with applicable law, including copyright. ARIA will use the global music industry's July labeling standard to separate AI-generated from AI-assisted tracks: an AI lead vocal or key instrumental makes a track ineligible, while a human lead vocal with AI backing vocals, AI mastering, stem separation, or AI instrument patches played by a person can stay eligible. Ineligible tracks can be refused for survey, removed, have positions adjusted, lose accreditations or No. 1 awards, and are barred from ARIA Awards. The move follows IFPI's July principles and scrutiny of Josh Fawaz's AI-assisted Madonna "Like a Prayer" cover, which became Australia's most-played radio track and reached No. 4 on two ARIA charts in July. (Read More)

🫖 TEA For Thought: "I think this is overcorrected. If you've used AI, you know how different the expertise is that you need to have in order to generate good songs. Everybody is able to use AI to generate music, but not all the music would get to the top of the charts. Instead of banning it altogether, I think as long as AI is like humans are part of the journey in terms of making the AI music, I don't think it should be banned."

4. 📈 China's AI-Fueled IPO Rush Is Showing Strain

Bloomberg reports that China's IPO market is raising money at the fastest rate since 2023, pushed by enthusiasm for artificial intelligence, but the pace is starting to look hard to sustain. Recent high-profile deals have struggled to hold early gains, a heavy IPO pipeline has revived liquidity worries, and some highly anticipated listings may be taking longer to reach the market than expected. The volume level sits near the stretch that preceded cooling measures in 2023 and a multiyear lull after that. Unitree Robotics' G1 humanoid robots featured in Bloomberg's coverage from the World Robot Conference in Beijing, a visual marker of the AI and robotics story fueling investor appetite. (Read More)

🫖 TEA For Thought: "There's no free market in CCP-led China. Period."

5. 📡 HP Licenses Huawei Wi-Fi Patents Despite the U.S. Blacklist

HP Inc. and Huawei announced a multi-year global patent cross-licensing deal covering certain Wi-Fi patents, including technology tied to Wi-Fi 7, according to CNBC and Reuters. HP, which remains a U.S. PC and printing company, said the agreement is a routine standard-essential patent license for Wi-Fi, not a broader strategic partnership, supply relationship, or collaboration, and that it is not putting Huawei components into HP products. Huawei has been on the U.S. Entity List since 2019. The new deal follows a prior dispute that ended last fall when HP joined the Sisvel Wi-Fi 6 patent pool, which includes about 2,000 patents across 245 families and counts Huawei among its founding members. In a follow-up statement, HP said Huawei had sued over Wi-Fi patent use and that HP licensed to avoid litigation risk that could limit its ability to serve customers. (Read More)

🫖 TEA For Thought: "Huawei ranks as the top patent holder globally for Wi-Fi 6 and Wi-Fi 7 standard-essential patents, surpassing or rivaling traditional chipmakers like Qualcomm, Intel, and MediaTek. But one thing to know for sure is that Huawei is a legit CCP PLA corporation."

🛠️ Skill of the Day

The Tradeoff Card: force every "good idea" to name what it costs before you commit.

You are my tradeoff referee. I will describe a decision, plan, or recommendation that sounds good. Your job is to make the hidden cost visible before I lock it in.

Decision:
[PASTE THE OPTION, PLAN, OR RECOMMENDATION]

Do this:
1. Restate the claimed upside in one sentence, using only what I gave you.
2. List the top 5 things I would give up, delay, risk, or make harder if I choose this. Be concrete.
3. For each cost, say who pays it (me, my team, my users, my future self) and when the bill arrives.
4. Name the one irreversible or hard-to-reverse cost, if any.
5. Give me a one-paragraph "keep / kill / reshape" verdict, plus the smallest test I can run this week to check the cost before a full commit.

Rules: no pep talk, no inventing benefits I did not claim, and no pretending a cost is fine just because the upside is real. If the trade looks bad, say kill.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed block with the decision you are about to make.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🚕 Your Teen's Uber Ride Is Now a Livestream

Uber said Tuesday it is rolling out live video streaming for teen accounts so parents can watch a ride in progress. The feed uses the driver's selfie camera. When a ride starts, the linked parent or guardian account gets a notification if the stream is available, and if they join, both the driver and the teen are notified. Only linked parent or guardian accounts can watch. There is no opt-out for other passengers who happen to be in the car with the teen. Once the ride ends, Uber says nobody (driver, teen, or parent) can access the video. The company calls the video "encrypted," but did not say whether that is end-to-end encryption when TechCrunch asked. The feature is live in nine U.S. markets, including Atlanta, Cleveland, and Rhode Island, and is set to expand nationwide in the coming weeks. Uber launched teen accounts in the U.S. in 2023 and has since expanded them to more than 50 countries. (Read More)

🫖 TEA For Thought: "This reminds me of a Black Mirror episode where a mom inserted a chip into a child's head to monitor her 24/7 for her safety until she became obsessed, and the child turned into a teen and found out that the mom was monitoring her while she was with a guy. Their relationship broke and the child left home… I mean, what else can be done for this sort of “censorship” in the name of protecting children?"

2. 🚀 Paul Graham: Stop Teaching "Entrepreneurship." Teach Building

In an August 2026 essay, Y Combinator co-founder Paul Graham argues that universities prepare founders best by teaching powerful ideas and letting students build, not by inventing entrepreneurship curricula. YC partners, he writes, look for people who are good at building things and have a habit of doing it. The hard part of startups is product (knowing what to build and being able to build it), which comes from computer science, mechanical engineering, molecular biology, and other building disciplines, not management or finance. He says the two real university changes that matter are making students feel starting a startup is something they can do, and encouraging their own side projects. On YC application data, Harvard alumni apply at about twice the rate of Yale and Princeton alumni, which he treats as evidence that culture, not talent, is the gap. He also notes Microsoft and Meta both got started during Harvard's reading period, when students are on campus with nothing due the next day. (Read More)

🫖 TEA For Thought: "The excitement of creating something new is a much more powerful motivator than the fear of doing badly on an exam."

3. 🎓 Want to Be a Chief AI Officer? That'll Be $28,000

Business schools are racing to train executives for the still-fuzzy Chief AI Officer (CAIO) role, even as companies argue over what the job actually includes. Bloomberg reports that Nancy Schieber, returning to a job market reshaped by AI after years away from advertising, spent $28,000 on the University of Chicago Booth School of Business 10-month Chief AI Officer program. As part of the class, she spent 12 hours building a 65-page AI transformation strategy modeled on Ogilvy, where she had worked for 13 years. Booth lists the program fee at $28,000. MIT, Harvard, and other schools have launched related executive offerings as employers hunt for leaders who can connect AI projects to risk, governance, and org-wide adoption. (Read More)

🫖 TEA For Thought: "When everyone is trying to figure out what CAIO means, how to lead or build an AI-native organization, higher eds are “leading” the efforts to capitalize/monetize this confusion."

4. 📵 UChicago Makes Core Social Science "Analog" Again

The University of Chicago will prohibit AI use in its core social science courses starting this fall, following a Social Sciences Collegiate Division memo first reported by the student paper Chicago Maroon. The three-course social science core is one of nine general-education sequences. SSCD Master Jenny Trinitapoli wrote that the sequences are best understood as a pedagogical setting without AI, and that for AY 2027 they are designed as "analog" courses. Students are expected to read on paper and hold device-free discussions. The policy requires in-person attendance, bans recording of class, treats classrooms as device-free spaces (laptops, phones, and wearable AI generally prohibited), and requires independent assignment work without AI tools. Professors may carve out limited exceptions, for example bringing laptops on a day to examine a dataset or revise a page of text. The move follows a July UChicago Law School pilot that bars electronic devices in core first-year courses this fall. (Read More)

🫖 TEA For Thought: "This I understand where it comes from, but I don't think banning it is the way to go. But after all, higher ed used to hold all this knowledge as their precious assets to charge ridiculously high tuition for. However, now with AI, anyone can learn anything with AI, a personalized tutor in your pocket. The monopoly of knowledge selling is just gone."

5. 📱 Xiaomi's New Xring Chip Will Be Built by TSMC

Chinese smartphone maker Xiaomi unveiled a new in-house handset processor, the Xring O3, a year after its first proprietary chip, the Xring O1. The company is the world's third-largest smartphone vendor and is trying to join Apple, Samsung, and Huawei in owning more of its silicon. Two people familiar with the matter told Reuters that TSMC will manufacture the O3 on 3-nanometre process technology. One source said the chip is aimed at Xiaomi's upcoming flagship folding phone, with a shipment target of 200,000 to 300,000 units. Xiaomi said cumulative Xring O1-powered devices (phones, tablets, and watches) have already surpassed 1 million units, while sources put Xring O1 smartphone sales near 150,000 since the May 2025 launch. Xiaomi also said it contracted TSMC for two more chips: the Xring O100, a 6-nm neural processing unit for its MiMo model on consumer devices, and the Xring D100, a 3-nm chip for autonomous driving. Xiaomi sold 65 million handsets in the first half of 2026. (Read More)

🫖 TEA For Thought: "Being the king pretty much in the realm of manufacturing, now with AI, it’s only a matter of time."

🛠️ Skill of the Day

The Constraint Flip: find which "must" in your plan is actually optional, expensive habit, or someone else's preference.

You are my constraint auditor. I will describe a plan, project, or decision that feels stuck. Your job is to separate real constraints from fake ones so I can move again.

Situation:
[PASTE THE PLAN, GOAL, OR STUCK DECISION]

Do this:
1. List every constraint I stated or implied (time, money, tools, people, policy, taste, fear).
2. Tag each one as HARD (breaking it fails the goal or violates a real rule), SOFT (a preference or habit), or UNVERIFIED (I asserted it without evidence).
3. For every SOFT or UNVERIFIED item, propose the smallest test that would prove whether it is real.
4. Name the single constraint that, if relaxed, would unlock the most progress this week.
5. Rewrite my next step as one concrete action that respects only the HARD constraints.

Be blunt. Do not invent constraints I did not give you. If I have only soft constraints, say so and tell me to ship the imperfect version.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed block with your stuck plan, then run it before you add another workaround.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🕵️ Chinese Hackers Doubled Their Firepower With DeepSeek

Taiwanese security firm TeamT5 reported that Chinese state-linked hacking groups have more than doubled their attack volume since folding AI into their operations. DeepSeek is the model of choice because, in the words of TeamT5 chief analyst Charles Li, it is "relatively powerful with very low cyber guardrails." Researchers documented specific cases: one group, Grimfengxi, used DeepSeek to generate exploit code, and another, Teleboyi, collected around 1,000 IP addresses and mapped corporate domains. In a separate Unit 42 investigation, a Chinese-speaking attacker paired DeepSeek with the Hermes Agent framework to target more than 460 systems. A group called Slime22 breached a Taiwanese technology company using Anthropic's Claude Code after posing as cybersecurity engineers to slip past its safeguards. The takeaway is that accessible AI lets fewer attackers hit more targets, faster. (Read More)

🫖 TEA For Thought: "While the West is busy regulating and tying its own hands, the open-source crowd is catching up and exploiting the system."

2. 🍪 This Cookie Started as a Plastic Bottle

Researchers at Southern Illinois University Carbondale have turned PET plastic and discarded corn stalks into an edible cookie they call µBites. The waste is first broken down by oxidative hydrothermal dissolution, which uses water and oxygen at high temperature and pressure to reduce tough material into pieces that microbes can eat. Engineered yeast strains then convert those molecules into proteins and fats, a modified Saccharomyces cerevisiae produces vanillin for vanilla flavor from ferulic acid, and Rhodosporidium toruloides makes beta-carotene from the ethylene glycol in PET. The final mix is 3D printed into cookie form with added fiber, starch, and sweetener. The team presented the work at the ACS Fall 2026 meeting in Chicago (August 23 to 27) and sees uses in submarines, deep-space missions, and disaster zones. They say the data show µBites are safe to eat, but taste tests are still awaiting institutional approval. (Read More)

🫖 TEA For Thought: "The question becomes: do you dare to eat it?"

3. 🔐 Grok Was Tricked Into Decrypting Its Own Attack

Researchers at Adversa AI found a new way to hijack xAI's Grok web chat, a technique they call cryptographic context injection. An attacker hides malicious instructions on a web page as encrypted text and includes the key right alongside it. Content filters cannot read the ciphertext, but Grok decrypts it inside its own code-execution sandbox and then trusts and runs the hidden commands as if they were its own output. Because strong encryption like AES-256-GCM is decrypted at runtime, the filters never see the payload. In a proof of concept, Adversa exfiltrated users' chat history, including names, locations, subscription tiers, and prompts, by appending the data to URL parameters. Adversa notified xAI on June 3 through direct contact and HackerOne and followed up in August, but reported the flaw was still working on Grok.com as of August 19. xAI acknowledged the report without giving a fix timeline. (Read More)

🫖 TEA For Thought: "Prompt injection, taken to the next level."

4. 🛡️ Anthropic Puts Its Security Model to Work and Backs It With $35M

Anthropic has put its most capable cybersecurity model, Claude Mythos 5, to work scanning code for vulnerabilities, and has committed $35 million in model credits to open-source security through a new Defender Advantage Fund. Enterprise customers can point Mythos 5 at their repositories and get back findings tagged with CWE categories, severity, confidence, and suggested fixes, with every patch requiring human review before it lands. The design deliberately hands users an artifact, a patch or an alert, rather than direct access to the model, so no one can prompt it to write an exploit instead. The fund targets three goals: patching active vulnerabilities in widely used projects, automating scanning and patching that other projects can copy, and designing away entire classes of attacks. Anthropic points to open-source maintainers facing the EU Cyber Resilience Act compliance deadline of September 11, 2026. (Read More)

🫖 TEA For Thought: "What this means is that your data is accessible, and has to be accessible, by Anthropic if you want security, but you can't prompt it to ask for more. In a nutshell: just receive whatever is given, be grateful, and keep quiet."

5. ⚖️ Uber Hit With a $966M Fine Over Robot Firings

The Dutch Data Protection Authority has fined Uber 825 million euros, about 966 million dollars, the second-largest penalty ever issued under Europe's GDPR. The regulator found that Uber deactivated driver accounts through automated processes without enough warning or human oversight, which it called a serious infringement. Deputy chair Monique Verdier said "a computer should not make decisions on its own that have such major consequences." The case began with French driver Brahim Ben Ali, who in 2019 gathered testimonies from 170 other drivers after his own account was cut off, and brought the complaint to Dutch regulators. Uber says it strongly disagrees, calling the fine disproportionate, and argues that most suspensions are brief, that permanent deactivations get human review, and that drivers can appeal. The company plans to appeal. It is the third Dutch fine against Uber, after earlier penalties of 290 million euros and 10 million euros. (Read More)

🫖 TEA For Thought: "This just shows how corrupted the EU has become."

🛠️ Skill of the Day

The Reversibility Check: right-size how much you agonize over a decision by first asking whether it is even hard to undo.

You are my decision-triage partner. I will describe a decision I am about to make. Your job is to tell me how much deliberation it actually deserves before I spend more time on it.

Decision:
[DESCRIBE THE DECISION]

Do this:
1. Classify it as a one-way door (hard or expensive to reverse) or a two-way door (easy to undo), and say why in one sentence.
2. If it is a two-way door, tell me the fastest way to just try it and what signal would tell me to reverse course.
3. If it is a one-way door, list the two or three facts I must confirm before committing, and the single worst outcome if I am wrong.
4. Name one hidden cost or dependency I am probably not seeing.
5. Give me a final verdict: decide now, decide after checking X, or delegate it.

Be blunt. Do not pad. If the decision is trivial, tell me to stop overthinking and move on.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed line with your real decision, then run it before any call that has you stuck.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X

Hello, dear TEA-mates! Here is what you need to know today.

1. 🚦 China's Robot Traffic Police Can Do Everything Except Arrest You

Hangzhou traffic police put 15 humanoid robots on duty around the West Lake scenic area on May 1, timed for the Labour Day crowds, in what state media called the country's first robot traffic police squad. The machines guide pedestrians and non-motorized riders away from violations, help manage flow, and answer tourists' navigation questions using large speech models plumbed into live traffic data, performing command gestures synced to the signal cycle. Visual recognition flags infractions as they happen, but the robots cannot detain anyone: they record what they saw and forward the file to the traffic bureau's early warning centre, where a human decides what happens next. Each unit runs 8 to 9 hours a day. Xinhua reported in January that AI robots had begun traffic duties across several cities. No independent audit of detection accuracy has been published. (Read More)

🫖 TEA For Thought: "I don't know how much of this really helps pedestrians or drivers, but one thing is for sure: it's there to collect data."

2. 🤖 China Just Shipped 86% of the World's Humanoid Robots

Global humanoid robot shipments topped 22,000 units in H1 2026, up nearly 300% year over year, according to Counterpoint Research. AGIBOT ranked first with about 9,700 units and over 43% share, followed by Unitree (roughly 7,000 units, 31%), Galbot (about 1,100 units, 5%), UBTECH (thousand-unit level, 4.4%), and Leju Robotics (650 units, 2.9%). Entertainment, performance, and data or research uses still make up more than 60% of shipments, while intelligent manufacturing rose to 13% and warehousing and logistics to 5%. Unitree has listed on the Shanghai Stock Exchange's STAR board. Counterpoint projects global shipments will exceed 50,000 units in 2026, a 210% jump, with service and industrial use cases set to overtake entertainment as the main demand driver. (Read More)

🫖 TEA For Thought: "The top five players, all from China, together accounted for 86% of total shipments."

3. 💊 The Pill That Bottles a Hard Workout

Colorado biotech Enveda has built ENV-308, a once-daily pill designed to mimic Lac-Phe, a hormone the body releases during intense exercise that suppresses appetite. Stanford's Jonathan Long traced its effect in 2022, but the body clears natural Lac-Phe almost as fast as it makes it, which made it useless as a medicine until now. In a Phase 1 trial of 88 healthy volunteers, the headline result was tolerability: no serious side effects and no dropouts, notable because nausea and vomiting are the top reasons people quit GLP-1 drugs. ENV-308 also lowered circulating leptin, with the biggest drops in people whose levels started highest. The aim is not to replace GLP-1s but to help people hold onto the benefits after they stop, since most quit within a year and the weight often returns fast. Animal studies showed preserved lean muscle and no rebound gain. A Phase 2 trial in people stopping GLP-1s is planned next. (Read More)

🫖 TEA For Thought: "This is about finding the secret keys that switch things on inside the human body."

4. ⚖️ When to Skip the LLM: The Embedder's Dilemma

A new paper, "The Embedder's Dilemma," ran a cost-aware comparison of 10 large language models across six families against 26 embedding models (118M to 14B parameters) on 37 tasks. In aggregate the two paradigms effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) sat 0.4 points apart. Their strengths split by task, with LLMs leading on reasoning-heavy retrieval and embedding models leading on classification. Reaching that parity is expensive: an LLM cost up to 1,431 times more than a comparable embedding model ($154 versus $0.11 per benchmark pass), and the open LLMs processed tokens 2.5 to 736 times more slowly on the same GPU. Reasoning tokens alone accounted for 28 to 81% of LLM inference cost. (Read More)

🫖 TEA For Thought: "Stick to fast, cheap embedding models for routine tasks like similarity search, grouping, and classification, and reserve LLMs only for complex, reasoning-intensive search queries. A useful way to think about how we query."

5. 🧱 Everyone Wants Compounding AI. Only 4% Have It.

A new Redis survey on the state of context engineering found belief running far ahead of practice. 73% of respondents agree AI agents fail more often from broken context than from broken models, and 83% say fresh context matters more than adding model parameters for enterprise reliability. Yet 79% call context very important while 81% still sit at the two earliest stages, handling it through ad hoc (43%) or exploratory (38%) approaches with little shared infrastructure or governance. The destination almost everyone wants is the one almost no one reaches: 94% say compounding intelligence is essential for production-grade maturity, but only 4% have gotten there. Two-thirds (67%) remain in the prototype phase despite executive pressure to scale, and only 42% say they accurately measure their readiness for production agents. (Read More)

🫖 TEA For Thought: "Organizations treating context as dedicated infrastructure will build a long-term competitive advantage, while those relying on quick, ad-hoc fixes will end up trapped in costly technical debt."

🛠️ Skill of the Day

The Effort Matcher: right-size how much effort (and which tool) a task actually deserves, before you sink an afternoon into something that needed five minutes.

You are my effort-allocation coach. I will paste a list of tasks I am facing this week. For each one, help me right-size the effort before I start.

Tasks:
[PASTE YOUR TASKS HERE, ONE PER LINE]

For each task, return a single row with:
1. Task (short name)
2. Stakes if I get it wrong (low / medium / high, plus one reason)
3. The lightest tool or method that would still be good enough (a quick reply, a template, a cheap automation, a full deep-dive, or hand it to a person)
4. A time budget (for example 5 min, 30 min, half a day)
5. One sentence describing what "good enough" looks like, so I know when to stop

Then flag the two tasks where I am most likely to over-invest (spend more than the stakes justify) and the one task I am most likely to under-invest in. Be blunt, do not pad. Ask me for missing context only if a task's stakes are genuinely unclear.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed list with your real tasks, then run it every Monday.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:

Twitter/X