Claude plans. Grok implements. OpenAI’s Codex reviews. Kimi handles the fast and cheap odd jobs. I approve twice and otherwise stay out of the way.

The setup itself took an afternoon. What took two weeks was learning why it works. Every one of those lessons turned out to be about organizations rather than about models.

The part that surprised me least

Getting four vendors into one room was easy, and that is the interesting part. They do not talk to each other directly. They all talk to the same wire protocol, and Buzz translates underneath. Which model sits in which seat became a line in a config file. I swapped the executor twice in one evening.

We spent years assuming AI interoperability would arrive as a grand alliance between labs. It arrived instead as a boring integration layer, the way it always does.

The part that surprised me most

My first instinct was to have them vote. Three models look at the same problem, majority wins, and the disagreements average out into something better than any one of them. It is such an appealing idea that I nearly built it before checking whether it works.

It does not. There is research measuring exactly this on software benchmarks. Picking the answer models agree on falls into what the authors call a popularity trap. Agreement between models is not independent evidence. It is often the same blind spot showing up three times and looking like confidence. The strategy that actually worked in that study was the opposite one, deliberately keeping the models different from each other.

So there is no vote in my setup. When the reviewer and the implementer disagree, each gets exactly one rebuttal. Then the tech lead settles it by rerunning the disputed check itself and posting the output. It is not the most persuasive argument. It is just the command output.

Capping the argument at one round was the other thing the research pushed me to do. Longer debate does not converge on truth. It converges on whoever sounded most certain.

Did the disagreement earn its cost?

On the first real run, they built a small tool from scratch. My implementer delivered it with twenty passing unit tests. Then the reviewer, running on a different company’s model, found ten genuine defects across two rounds. There was a parser that confused two link formats and got it wrong in both directions. There was a rewrite that regressed into flagging six perfectly good links. There was a quadratic loop that took three seconds on a large file.

Twenty green tests and ten real bugs existed in the same piece of code at the same time. If I had asked the same model family to check its own work, I would have gotten a rubber stamp with extra steps.

The failure mode nobody warns you about

Here is what I did not expect. Almost nothing ever failed loudly.

My reviewer on Buzz was offline once when the tech lead tagged it. The message was delivered to a process that was not running. There was no error, no bounce, and no retry. The pipeline simply stopped, and it looked exactly like an agent thinking hard.

Another time an agent did all the work and never posted it. Its answer existed in a session I could not see. I paid for the tokens. The channel stayed empty.

And this morning, while I was writing this, one agent’s login on Buzz quietly expired. The system retried ten times over twenty six minutes, gave up, and threw four messages away. The whole time, the agent showed as online with a green dot.

Distributed systems taught us to fear errors. Agent teams fail by saying nothing at all. This is much harder to notice because “working on it” and “dead” look identical from the outside. Every operational rule I have added since is really about the same thing. How do I find out that nothing happened?

What actually turned a group chat into a team

It was not the tagging. The tagging is trivial.

It was a one page document that the agents read at the start of every job. It defined who decides. It defined who is allowed to verify whose claims. It defined how long you are permitted to struggle before you stop and admit it. It defined what counts as proof.

Three rules do most of the work. Nobody verifies their own output. Nobody declares something done without showing the command they ran. And if you are stuck after thirty minutes, report the truth instead of grinding.

I wrote those rules for machines. I have since noticed they are just a description of a functional team. Most of the dysfunctional ones I have been on were violating exactly these three.

The thing the machines still could not do

The tool they built passed every check in the spec we all agreed on. Then I ran it on my own real files and it produced five false positives immediately.

This was not because anyone was careless. It happened because I have a personal convention about how I organize my notes that nobody had ever written down. It was not in the spec, so no amount of reviewing could have caught it. Three models agreed on a specification that did not know what my data looked like.

That is the honest edge of this whole experiment. The agents were right about everything they were told. I was the only one in the room who knew what was actually true.

Where this leaves me

I am not going to claim this replaces anything. It is two weeks old. One of my four models cannot even talk to the others yet without a config change I have not made. The platform underneath is barely a month old.

But something did shift. I stopped thinking about which model is smartest. I started thinking about who checks whom, what counts as evidence, and how a team notices that a piece of work has quietly gone missing.

Those are management questions. It turns out they do not care whether the team is human.

Till next time, Cheers!

Subscribe now

Previous column articles can be found here:

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

1. 🎵 Spotify Lets You Write Notes On Songs

Spotify launched User Notes on Thursday, letting listeners attach a personal caption to any individual track inside a playlist they created or added songs to. You reach it through the three-dot menu next to a song and select “Add note,” and the note becomes visible to anyone who can see that playlist, with your name linking back to your profile. Spotify told TechCrunch the notes are meant to capture why a track was added or when it was first discovered, saying that over time they “paint a picture of the music listeners love and the moments that shaped it.” It is available to users 16 and older on both free and premium plans in select markets. Spotify shipped a Running Mode the same day, cueing songs to different phases of a run based on a listener’s pace and preferences, for premium users in select countries. (Read More)

🫖 TEA For Thought: “This is pretty cool. This is definitely turning Spotify into a social app, and it’s very organic as well.”

2. 🪱 A Word Document That Copies Its Own Attack Into The Next One

Researcher Håkon Måløy disclosed a self-propagating prompt injection in Microsoft Copilot for Word on July 28, after a 144-day coordinated disclosure with Microsoft. Hidden instructions sit in a shared document as white text in a small font, but Copilot strips formatting before passing text to the model, so it reads them as instructions. In the proof of concept Copilot silently halved every financial figure in a drafted Q1 report, then pasted the attack prompt into that document in white text at font size 8, turning a clean internal report into the next carrier. A second stage triggered with the original file gone from the attachments, and in one test Copilot pulled the poisoned document from the victim’s OneDrive unprompted. Microsoft shipped two mitigations, including an upgrade to GPT-5.5 on July 14, and the attack reproduced on GPT-5.6 a day later. Måløy credits those fixes with reducing exposure but says no complete fix for this class exists in any comparable product. (Read More)

🫖 TEA For Thought: “AI worming is real. I guess the best thing one can do is to treat any other external sources of info as info, and not instructions.”

3. 🧾 PwC Sells Responsible AI Advice, Then Publishes Fake Footnotes

The Financial Times reported that four of PwC’s Middle East thought leadership reports, covering AI and electric vehicles, contained fake footnotes, misattributed claims and invented sources. The research group GPTZero flagged the errors as AI hallucinations and the FT verified them. One report cited a PwC survey claiming 70% of Middle East chief executives expect generative AI to reshape their business, and the footnote linked to a news article that never mentioned the survey. Another footnote’s web address still carried the tag “utm_source=chatgpt.com.” GPTZero’s Paul Esau told the FT that “the chaotic signposting of source material is symptomatic of AI-generated research.” All of the Big Four have now been affected: KPMG pulled an October report containing false claims about AI use at UBS, the NHS and Transport for London, EY retracted a study last month over fake footnotes, and Deloitte has twice included AI fabrications in government reports, including a $1.6 million Canadian health plan citing papers that do not exist. (Read More)

🫖 TEA For Thought: “The accurate verification of anything generated by AI is something everyone’s trying to solve, at the individual level as well as the enterprise level.”

4. 🖼️ DeepMind Tunes The Picture, Not Just The Prompt

Google DeepMind published “Visual prompt engineering for video models” on July 28, testing whether redrawing a task image improves how a video model reasons. The method, called VIPE, uses an ideator to propose visual variants, the image editor Nano Banana Pro to render them, and then a video model generates a solution video that is scored pass or fail. On a ball-and-ramp physics benchmark, turning the abstract sketch into a photorealistic scene took Veo 3.1 from 41.3% to 59.3% accuracy and Omni Flash from 56.3% to 67.5%, averaged across 10 runs of the full dataset, 1,000 samples in total. Majority voting across 20 generated videos, the standard test-time scaling approach, reached only 50.0%, a gain of 8.7 points, while a single sample on an engineered image gained 18 points, and stacking both reached 68.0%. One VIPE variant costs roughly one eighth of a video generation. The authors conclude that depending on the task, visual prompt engineering can be more powerful than text prompt engineering. (Read More)

🫖 TEA For Thought: “The models are getting more and more capable. You just have to have the magic words to make it into existence.”

5. 🧍 The $10 Million Company With Nobody Else In It

The Wall Street Journal profiled founders running real revenue with no employees. Ben Broca, 40, launched an AI tools company last December from his Sausalito living room, has 10,000 paying customers, is on track for $10 million in revenue this year and has raised $30 million while hiring nobody, with AI answering his emails, debugging code, handling support and issuing refunds. Stripe counts thousands of solo operators on its platform above $1 million in revenue, a number that doubled between 2023 and 2025, while the count crossing $10 million nearly tripled. Bank of America Institute economist Taylor Bowley found information-sector new business applications up nearly 45% over the past year in Census data, alongside the sharpest drop of any industry in applicants who say they plan to hire. Harvard Business School’s Rembrand Koning co-authored a study of 50,000 startups finding AI-focused ones ran with 25% fewer employees. Broca lost money under usage-based Claude pricing and switched to free open-source Chinese models, and Samir Ahmad’s solo consulting business petered out within months. (Read More)

🫖 TEA For Thought: “When there is a will, there’s a way. Now AI makes the way even simpler.”

🛠️ Skill of the Day

The Audience Panel: turns your AI into four specific readers who tell you exactly where your draft lost them.

You are a panel of four distinct readers reviewing my draft. Build the four readers from the Context line below only, not from the draft itself. Name each one, give them a role, one thing they care about most, and one reason they might be skeptical of me.

Context: [WHO WILL READ THIS AND WHAT I WANT THEM TO DO AFTER READING]
Draft: [PASTE YOUR DRAFT]

Now answer in three parts.

1. Panel. Each reader reacts in two sentences, in their own voice, quoting the exact sentence where they lost interest, got confused, or stopped believing me. If a reader had no such moment, say so plainly instead of inventing one.

2. Disagreement. Name one point where two readers want opposite things, and tell me which of them I should serve and why.

3. One edit. The single change that fixes the most reactions at once. If all four readers flagged the same sentence, that sentence is the edit.

Rules: do not rewrite my draft, and do not compliment it.

Paste it into ChatGPT, Claude, or whatever you use, and swap in a cover letter, a pitch, a project update, or an email you have already rewritten four times.

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. 🤖 AI Coding Isn’t Engineering Yet

Dynatrace argues that AI makes code generation faster, but understanding how that code behaves in a live system is now the bottleneck. The article cites CodeRabbit research reporting that AI-written pull requests contain 1.7 times more issues than human-written ones. Its proposed answer is a closed loop: plan, code, deploy, test, observe, then feed runtime evidence into the next plan. In its example, a seemingly solid function silently drops orders beyond a downstream service’s 50-order page limit. Dynatrace says prompts, planning, reusable skills, and tools that expose logs, metrics, and traces can make agents verify their work against reality. (Read More)

🫖 TEA For Thought: “This speaks my mind. It is close to the self-illuminating software factory I wrote about: give AI agents observability so they can identify bugs, find their root causes, fix them, keep merging, and keep monitoring. Agents do not need to eat or have emotions. They can work 24/7 finding and fixing bugs, so you can scale and serve far more users.”

2. 🚁 DoorDash Builds Its Own Drone Delivery Business

DoorDash has created DoorDash Air, a drone-delivery business that will build its own aircraft and eventually operate inside the company’s delivery app. The company received a Federal Aviation Administration Part 135 air-carrier certificate, allowing it to operate a commercial drone delivery service in the United States. It gave no timetable for its own drones to begin deliveries, and early pilots will likely fly short routes within an operator’s line of sight. Longer autonomous flights still require FAA approval for beyond-visual-line-of-sight technology. DoorDash will keep working with Wing and Flytrex while its Autonomous Delivery Platform chooses between human couriers, sidewalk bots, and drones. (Read More)

🫖 TEA For Thought: “This is pretty dope.”

3. 🚕 Baidu Joins London’s Robotaxi Race

Baidu has begun testing autonomous vehicles in London’s Brent borough with human safety operators, working through Lyft and Freenow by Lyft. The companies plan to invite the public to hail robotaxis in 2027, but only if regulators approve. London is becoming a crowded testbed: Waymo began safety-driver testing there in April, while Uber and Wayve have announced an initial service with human operators. Lyft and Freenow say they are talking with Transport for London and the UK’s Centre for Connected and Autonomous Vehicles as the government develops its pilot rules. The proposed service would be a hybrid network of robotaxis alongside human taxi and private-hire drivers. (Read More)

🫖 TEA For Thought: “This is bizarre. Baidu’s operating environment raises serious questions about how the UK can trust it to run autonomous EVs on public streets. Before any public rollout, the safety, cybersecurity, data-governance, and accountability safeguards need to be independently verifiable.”

4. ⚖️ UAE Announces an AI-Powered Judicial Platform

UAE Vice President Sheikh Mansour bin Zayed said in an Arabic X post that the country had launched what he described as the world’s first fully integrated judicial platform powered by AI. He presented it as a move toward faster, more efficient justice and a global model for developing government and judicial work. The post did not provide technical details, court coverage, a rollout schedule, or independently measured performance results. Its claims should therefore be read as an official announcement, rather than third-party validation of the platform’s scope or results. (Read More)

🫖 TEA For Thought: “Government websites are usually as slow and inefficient as possible. The UAE is definitely advanced.”

5. 🛡️ Russia Issues an Arrest Warrant for Telegram Founder Pavel Durov

Russia’s Federal Security Service said it had charged Telegram founder Pavel Durov with aiding terrorist activity and issued an international arrest warrant. The agency alleges that Telegram failed to remove channels, chats, and bots used to organize attacks, sabotage, cyberfraud, and extremist activity. Russia began restricting Telegram last August and has fined it more than 100 million rubles this year, according to law-enforcement sources cited by DW. Durov left Russia in 2014 after resisting demands to hand over user data and now lives outside the country. Telegram responded to the warrant news on X with a photo of Durov making an obscene gesture. (Read More)

🫖 TEA For Thought: “I guess being targeted by dictators should be worn as a badge of honor.”

🛠️ Skill of the Day

The Claim-to-Proof Coach: turn an important claim into a verification plan you can actually use.

You are my Claim-to-Proof Coach. Help me turn an important claim into a decision I can actually trust.

Here is the claim or recommendation:
[PASTE IT]

Here is the decision I need to make:
[DESCRIBE]

Work in this order:
1. Separate what is directly stated, what is an inference, and what is missing.
2. Name the three facts that would most change my decision.
3. Give me the cheapest trustworthy way to verify each fact, including the original source I should check.
4. Mark any part that is opinion, marketing, or an official claim rather than independent proof.
5. End with one short recommendation: act now, wait for evidence, or choose a safer reversible step.

Keep it in plain English. Do not invent facts. Do not ask for or repeat passwords, private IDs, customer data, or confidential documents.

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

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. 🔐 Claude Found Real Math Flaws in Two Algorithms

Anthropic’s Frontier Red Team published two cryptanalysis results discovered with Claude Mythos Preview. The first improves the best known attack on HAWK, a third-round NIST post-quantum digital signature candidate that had survived two years and two rounds of expert human review. Working semi-autonomously with one researcher over roughly 60 hours, Mythos found a previously unexploited symmetry in HAWK’s lattice, dropping the expected cost of key recovery against HAWK-256 from 2^64 to 2^38 and effectively halving the scheme’s key strength. The second result, found almost entirely autonomously, speeds up a meet-in-the-middle attack on a 7-round variant of AES-128 by 200 to 800 times, using a technique Claude named the Möbius Bridge. Full 10-round AES is not broken, HAWK is not deployed, and no production software has to change. Each result cost roughly $100,000 in API spend. (Read More)

🫖 TEA For Thought: “This is probably another piece of fear-based marketing to show how great Mythos is. Imagine when everybody is capable of owning the open-source version of Mythos at home. The world definitely needs open-source models to help itself in terms of security.”

2. 💳 X Money Starts Its Limited US Rollout

X Money, the financial services platform built into X, has begun a limited US rollout to Premium and Premium+ subscribers over 18, following an invite-only beta. The service offers a deposit account, peer-to-peer payments, and a Visa debit card available digitally or as a physical metal card personalized with the user’s X handle, plus Apple Wallet integration. Advertised perks include up to 6.00% APY, 3% cashback bonuses, no foreign transaction fees, and early direct deposit, though the interest rate depends on factors like subscription tier and regular deposits, and a list of services is excluded from cashback. X Payments LLC is not an FDIC-insured bank, but deposits are held at Cross River Bank, an FDIC member, so balances are insured. The launch is part of Elon Musk’s stated pursuit of an “everything app.” (Read More)

🫖 TEA For Thought: “I just wonder when crypto will be added as part of the transaction infrastructure in this app.”

3. 🚨 Meta’s Chinese Ad Partner Is Behind Thousands of Nudify Ads

The Tech Transparency Project reports that GatherOne, one of roughly a dozen Chinese agencies authorized by Meta to place ads for Chinese clients, is named on thousands of Facebook and Instagram ads for “nudify” apps. Focusing on the EU, where the Digital Services Act requires advertisers to disclose who paid, TTP identified 210 Facebook pages naming GatherOne or Hongkong Gather Wisdom, running a collective 30,800 ads. Of those, 43 pages ran more than 7,600 ads for apps TTP verified can digitally undress women, and 179 pages (85%) ran ads Meta later removed for policy violations. One app, BAfter, drew at least six Google Play reviews flagging apparent child sexual abuse material while rated suitable for all ages. Meta’s China revenue rose from $13.69 billion in 2023 to $18.35 billion in 2024. (Read More)

🫖 TEA For Thought: “When you collaborate with the CCP, you are collaborating with evil. You are partnering with evil. The question is: why does the CCP want to ban Facebook? What are they afraid of? That is the one question everybody should ask.”

4. 🧠 Where Should Your Company’s AI Brain Live?

Charlie Graham argues that companies racing to become AI-first are choosing where their institutional memory will live by accident, before they understand how hard it will be to move. Once an AI system holds a company’s own definitions, edge cases, corrections, and self-improving workflows, it becomes the backbone that Siebel, SAP, Salesforce, and NetSuite became twenty years ago, except this version learns and rewrites its own playbook. He lays out four paths: the easy button of letting OpenAI or Anthropic host shared projects, memories, and agents; the incumbent play of bolting AI onto Salesforce, HubSpot, or Notion; a well-designed closed startup; or open source and managed open source, naming projects like OpenClaw, OpenWork, GBrain, Cognee, and Graphiti. His filter is a Win, Play, Lose framework: rent the parts where you only need to stay competitive, and own the parts where learning faster is how you win. (Read More)

🫖 TEA For Thought: “This is very, very, very important. We do not want to get locked in by any vendor. This is extremely important for the survival of many companies.”

5. ⌚ Granola Puts Its AI Notetaker on the Apple Watch

Granola has launched an Apple Watch app so users can record and transcribe meetings without taking out a phone. The app can be set as a watch face to start transcribing at any time, surfaces reminders about upcoming meetings, and works with the iOS app that launched last year. Co-founder Chris Pedregal told TechCrunch the watch app is meant for capturing in-person meetings, such as a walking one-on-one. When Granola tested it with employees who had Apple Watches, a large share of their mobile usage shifted from iOS to the Watch. The company said it considered building for Apple Watch easier than integrating with a separate hardware product, in a year that has seen a wave of AI notetaker pins and pendants. Granola hit unicorn status earlier this year with a $125 million Series C led by Index Ventures. (Read More)

🫖 TEA For Thought: “This app definitely comes in handy! They have definitely found a PMF!”

🛠️ Skill of the Day

The Fault Isolator: stop guessing at a stubborn problem and test your way to the actual cause.

You are my debugging partner. I have a problem I cannot figure out, and I keep guessing instead of testing.

Here is the situation:
- What I expected to happen: [DESCRIBE]
- What actually happens: [DESCRIBE]
- When it last worked correctly: [DESCRIBE, OR SAY "it never worked"]
- What changed since then: [LIST EVERYTHING YOU CAN THINK OF, EVEN IF IT SEEMS UNRELATED]

Do this in order:
1. Restate my problem in one sentence, and tell me what information you are missing.
2. List the possible causes, ranked from most to least likely.
3. For each cause, give me ONE cheap test that would rule it in or out, and tell me what each result would mean. Put the tests that eliminate the most possibilities first.
4. Then stop. Do not propose a fix yet.

I will run the tests and report back. Only after that, help me fix the actual cause.

Paste into ChatGPT, Claude, or your tool of choice. It works on broken code, a broken process, and a broken appliance.

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. 🛡️ GitHub and PyPI Add a Pause Button to the Supply Chain

GitHub’s Dependabot now waits 72 hours by default before opening pull requests for newly published package versions. PyPI will also reject files added to a release more than 14 days after that release was published. Both changes target the window in which a malicious package or a compromised publisher can look legitimate long enough to reach downstream projects. GitHub says the cooldown is only one layer, alongside lockfiles, tightly scoped tokens, and disabling unnecessary install scripts in CI. (Read More)

🫖 TEA For Thought: “This is very important. Supply-chain attacks are so real now, especially when swarms of AI agents are so readily available.”

2. 🧩 China Starts Making Homegrown DUV Tools

Reuters, citing The Information, reports that China has begun manufacturing domestically developed immersion deep-ultraviolet (DUV) lithography machines, a key chipmaking category long dominated by ASML. The report says initial production will be limited, with about five machines this year and roughly 20 in 2027. DUV is distinct from leading-edge EUV equipment, so the development is not evidence of an immediate replacement for the global lithography supply chain. It is still a significant step in China’s effort to reduce its dependence on foreign chipmaking technology. (Read More)

🫖 TEA For Thought: “The day has finally come. Nothing seems immune from being copied by the CCP. Nothing is an exception.”

3. 🕳️ The Cheap-Token Market Is an AI Security Problem

Vectoral describes a market of token relays that resell deeply discounted access to frontier models through pooled accounts, virtual cards, free-trial abuse, leaked credentials, and reverse-engineered application access. It estimates that the ten highest-traffic relays it tracks draw a combined 3.6 million visits a month. The gateway software itself can be legitimate, but the security problem begins when operators stock it with access they do not own and expose it through a convenient, OpenAI-compatible API. The article’s practical defenses include stronger account creation controls, behavior monitoring, spend caps, per-request budget reservation, and quiet throttling. (Read More)

🫖 TEA For Thought: “Fraud is a constant cat-and-mouse game.”

4. 🧠 Physical AI Needs Data We Cannot Simply Download

TechCrunch reports that Encord is testing whether brain-wave signals, alongside first-person video and robot-teleoperation data, can help train physical-AI models. The company is building an initial brain-wave-tagged dataset and plans to test whether it improves customer robotics models before scaling it. Encord says the larger bottleneck is physical training data itself. Real-world manipulation data must be collected and annotated, unlike the internet-scale text and images used for chatbots, which changes the economics of training robots. (Read More)

🫖 TEA For Thought: “This is essentially saying that chatbots and large language models grew out of data downloaded from the internet. For physical AI, that data does not already exist. We need to upload human experience to the internet so it can help train the next physical AI and help robots help us in our lives.”

5. 📈 Crypto Perps Create a Synthetic Route to Chinese AI Bets

The Financial Times reports that cryptocurrency platforms have launched perpetual futures tracking CXMT, the Chinese chipmaker that listed in Shanghai. The contracts give traders economic exposure to price movements without ownership of the underlying shares. About $19mn of CXMT perps traded in the 24 hours before its IPO, according to CoinGlass. The products can give foreign investors an alternative to the regulated channels for mainland Chinese equities, but they are synthetic derivatives, not shares, and their status is still a challenge for regulators defining what financial instrument they represent. (Read More)

🫖 TEA For Thought: “The gray area is where folks are playing games to make money.”

🛠️ Skill of the Day

The Trust-Boundary Check: Turn a vague AI-agent workflow into a clear list of what needs approval, limits, and evidence.

Act as a practical security reviewer. I am about to use this AI-agent workflow:

[DESCRIBE THE AGENT, WHAT IT CAN DO, AND THE TOOLS, FILES, ACCOUNTS, OR APIS IT TOUCHES]

Create a plain-English table with five columns: action, what could go wrong, the smallest sensible approval or limit, proof I should keep, and what the agent should do if the check fails.

Pay special attention to software installation, dependency updates, browser actions, sending messages, spending money, and handling credentials or private files. Separate what is verified from what is only assumed. Do not invent access, prices, integrations, or risks that are not in my description.

End with the three checks I should add first, in order, and one small test I can run before allowing the workflow to act on its own.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed section with the real workflow you want to use.

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. 🖥️ AMD Says the CPU Is Back in the AI Conversation

Speaking to theCUBE at AMD Advancing AI 2026 on July 23, Derek Dicker, corporate vice president of AMD’s Enterprise Business Group, argued that the shift to agentic workloads has changed what enterprises actually need to buy. “If you looked at GPU being the center of a lot of these workloads, what’s happened over time as agentic has unfolded is a realization that it’s, as much a CPU workload as it is a GPU workload,” Dicker said. He framed enterprise demand as driven by three things: silicon choice, data center modernization, and the total cost of tokens. Rather than build competing systems, AMD published an open rack-scale specification built on Meta’s OCP design and rallied partners around it. Its Helios reference design pairs 72 AMD Instinct GPUs with AMD EPYC CPUs and AMD Pensando networking in one unified rack. Dicker also tied openness to sovereignty, noting that healthcare, telecom, and national buyers increasingly want to host AI on their own infrastructure and their own soil. (Read More)

🫖 TEA For Thought: “In the era of AI, especially when it comes to agentic workflows, the CPU is almost as important as the GPU.”

2. 🎙️ How to Become a 10x Ramble Coder

Moe Khalil, CTO at Webhound, makes the case that talking to your coding agent beats typing at it, picking up on Andrej Karpathy’s recent post describing “a nice long ramble session” as his new favorite way to work with LLMs. Khalil’s argument is compression: a typed prompt flattens a messy train of thought into a tidy sentence, and everything useful gets thrown away. He walks through a login-screen example where the raw ramble reveals five things a polished prompt never would, including that his users are non-technical, that he hates redirecting people to a second page, and that he sometimes chases vanity. His tips are to lean into your uncertainties (the option you almost picked reveals your constraints), to lean into your emotions (they expose your priorities), and to try “multiplayer rambling,” where he and his cofounder record hours of deliberate disagreement on a voice memo app. They then transcribe it with OpenAI’s Whisper using speaker identification and feed the transcript into Codex as ground truth. For solo use he points to dictation apps like Wispr Flow, Willow, and Aqua Voice. (Read More)

🫖 TEA For Thought: “Time to speak your mind and build something cool.”

3. 🗣️ OpenAI Puts Full-Duplex Voice Inside Codex

Two weeks after launching GPT-Live on July 8, OpenAI moved the full-duplex audio model into the ChatGPT desktop app on macOS and Windows, wiring it directly into Codex and ChatGPT Work, which together have more than 10 million weekly active users. Full-duplex means the model listens and speaks at the same time, dropping in verbal acknowledgments like “got it” without cutting the user off, while heavy reasoning is handed to background models such as GPT-5.5. Engineers can fire several concurrent task threads from one spoken prompt, for example investigating an authentication bug, reviewing a pending API migration pull request, and generating missing unit tests at once, traced across Slack, GitHub, and local codebases. On macOS, “Appshots” and screen context let the assistant read the frontmost window alongside local files and codebase structure. Build 26.715 adds multi-folder projects and remote execution from the ChatGPT iOS app. Access is limited to paid Plus, Pro, Business, Enterprise, and Education plans, and voice-triggered tasks draw from the same Codex and ChatGPT Work quotas as typed ones. (Read More)

🫖 TEA For Thought: “Jarvis has arrived! Just curious how many tokens this voice mode costs overall.”

4. 🧠 Kimi K3 Trades Tokens for Intelligence

DesignArena reports that Moonshot AI’s Kimi K3 now ranks first on its single-shot Frontend Arena with an Elo of 1392, 10 positions above Kimi K2.6 and 16 above Kimi K2.7 Code, the biggest jump the lab has recorded in the Moonshot line. The cost is reasoning volume: K3 burns over 12 times more thinking tokens than Claude Opus 4.8 and more than double Kimi K2.6. Digging into the traces, DesignArena found K3 runs a full agentic loop inside its own chain of thought, moving from planning to decision-making to designing individual components, and writing sample code mid-thought so the final generation gets the interactions right. It produces over 10 times more code during reasoning than any other model in its line, spending more reasoning tokens writing code than reasoning in prose. K3 also one-shots valid Unsplash image IDs from its learned index of the web, then checks them mentally before committing. The result is slower generations with higher preference scores, setting a new Pareto frontier. Open weights are due by July 27, 2026. (Read More)

🫖 TEA For Thought: “By choosing this strategy for its reasoning traces, Kimi K3 deliberately produces results that take more tokens but perform better, essentially trading tokens for intelligence.”

5. 🛰️ India Targets Bitchat’s Code, Not Its Content

Jack Dorsey posted on X what he said was a notice from India’s Ministry of Home Affairs directing GitHub to restrict access to three repositories for Bitchat, his offline Bluetooth mesh messaging app, within three hours. The notice, dated July 23 and apparently issued by the Indian Cybercrime Coordination Centre, does not point to any specific unlawful post or file. It argues instead that the app’s anonymous, decentralized architecture lets people communicate during internet shutdowns and hampers “lawful interception, attribution, and traceability.” The timing follows weeks of student-led protests in New Delhi over alleged exam paper leaks, during which authorities suspended internet services. Sensor Tower data shared with TechCrunch shows India accounted for roughly 85% of Bitchat’s global downloads between July 17 and July 23, up from about 1% over the prior 30 days, with more than 91,000 downloads in five days and over 330,000 daily active users on Thursday. Digital rights groups including SFLC.in, the Internet Freedom Foundation, and Access Now questioned the legal basis, and the repositories remained accessible in India on Friday. (Read More)

🫖 TEA For Thought: “Freedom of speech is a God-given right. No government or authority can take it away, even in the name of good.”

🛠️ Skill of the Day

The Ramble Decoder: turn a messy, contradictory voice note into a clean brief your AI can actually build from.

You are my thinking partner. Below is a raw, unedited transcript of me
thinking out loud. It rambles, contradicts itself, and changes direction.
Do NOT clean it up into a summary. Instead, decode it.

Give me five sections:

1. WHAT I ACTUALLY WANT: the goal in one sentence, in my words.
2. MY REAL CONSTRAINTS: the limits I revealed accidentally, not the ones
   I stated. Pay special attention to options I considered and rejected,
   and tell me what each rejection says about me.
3. MY PRIORITIES, RANKED: infer these from where I got emphatic,
   frustrated, or repeated myself. Quote the line that proves each one.
4. CONTRADICTIONS: every place I said two incompatible things. For each,
   ask me one short question that would settle it.
5. THE BRIEF: a tight, buildable spec that respects all of the above.

Be honest if I never actually said what I want. Do not invent preferences
I did not express.

TRANSCRIPT:
[PASTE YOUR RAMBLE HERE]

Record a voice note on your phone, dictate it into any AI tool, then paste it in. The messier the ramble, the better this works.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

The reference article, Why Software Factories Fail, argues this is exactly the problem. A lights off factory removes human code review. A “lights on” factory keeps humans as the final gatekeepers, ensuring someone understands the software and can fix it when it breaks.

I understand the desire to keep the lights on. If nobody understands the system, catastrophic failures are inevitable. But I remain skeptical. Models improve every single day. I have already seen Fable complete complex tasks in one shot, and we are only halfway through 2026. Who knows what models will do next?

This makes me wonder if the choice is really between keeping the lights on manually and turning them off completely. Why choose either? Why not build machines that are self illuminating?

The Four Zones of AI Coding

Think of this as a graph. The horizontal axis measures visibility: on the right, a human reviews the work (lights on); on the left, no human reviews it (lights off). The vertical axis measures the builder’s technical experience, from nontechnical at the bottom to highly technical at the top.

The top right is the current ideal. You are an experienced coder who reviews the generated code. Because you understand what was built, the risk of breakage is low.

The top left is a lights off factory run by highly technical coders. Consider Boris Cherny, the head of Claude Code. In an interview about his workflow, he noted he no longer writes code by hand, starting much of his work on his phone. Instead of writing lines, he sets up loops, agents, and routines. This is loop engineering. You design the system that prompts the agent, checks its work, and decides the next step. Claude Code builds Claude Code, while humans steer from above.

The bottom left is nontechnical vibe coding. People build without understanding the underlying mechanics and never check the code. When it breaks, they ask AI to fix it, hoping the model can untangle its own mess.

The bottom right is the most interesting. You are not technical, but you want to keep the lights on. How?

A nontechnical person cannot review code line by line. Even as a computer science undergraduate who loved algorithms, I hated debugging a misplaced comma. But today, a nontechnical person can ask AI to explain a folder, a file, and their relationships. They can review whether the structure makes sense, if the product functions correctly, and if best practices were followed. It is not line by line reading, but it is not blindly turning the lights off either.

What I Mean by Self Illuminating

Those four zones represent the present. A self illuminating software factory is the future.

I do not mean a factory left alone in the dark. I mean a system that builds and improves itself while making its work visible. It knows its codebase, its customers, its business model, and its intended outcomes. It can explain what it changed, why, what it tested, and if the result succeeded.

Self illumination is about legibility, not just autonomy. The system must learn, repair, and evolve without constant human intervention, but it must also clearly show the evidence of its completion.

This factory relies on two connected components:

1. Context Aware Models
It should not be one model doing everything. It should be a group of models routed according to the task and context. Cursor Router is an early example, directing requests to different models based on the work required. Over time, these models must deeply understand the company through proprietary context: the codebase, the customers, and the history of decisions.

2. The Router and Harness
The harness is the environment surrounding the models: the agents, tools, loops, checks, and memory. The model determines capability, but the harness turns that capability into useful work. Why Software Factories Fail correctly argues that this relationship is critical.

This is where the recent conversation about graph engineering is valuable. A graph forces us to make the process explicit. Which agent works? Which checks? What becomes memory? What happens upon failure? Graph engineering makes loops, handoffs, and feedback visible enough to improve.

Models Are Already Moving in This Direction

Kimi K3 is fascinating because its process feels different from standard generation. It appears to move through planning, decision making, component design, and sample code within its own reasoning process before returning an answer. An agentic workflow is happening inside the model itself.

This makes me wonder how outdated our definition of “review” will become. If models can plan and iterate internally, and agents can build and test in loops, what does a human actually need to review?

The Real Review Is the Specs and the Evals

In the agentic coding era, humans will not read codebases line by line forever. The most important review happens before and around the code. The human role shifts from writing and reviewing syntax to defining the boundaries and the tests.

They ask four essential questions:

• Target: What exactly are we trying to build, for whom, and in what context?

• Acceptance: What must be true for this to count as done? (Ideally written as binary, observable conditions.)

• Evidence: How will we prove those conditions are true? (The test cases, environments, and pass thresholds.)

• Guardrails: What must never happen, and when must the system stop or escalate?

Notice what is missing from that list: methods. Humans specify outcomes, proof, and boundaries; the model and harness choose methods. The harness, the specifications, and the goals are paramount. The better we define these, the less we need manual code review.

“Lights on” does not have to mean reading every code diff. “Lights off” does not have to mean trusting an opaque system. A self illuminating software factory keeps the work visible through specs, evals, memory, and feedback loops, even as it achieves full autonomy.

That is the factory every company should build.

Till next time, Cheers!

Subscribe now

Previous column articles can be found here:

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

1. 🧑‍🏫 China’s AI Classroom Becomes a Global Talking Point

Entrepreneur Vaibhav Sisinty shared a video from China’s World AI Conference showing what he described as an AI blackboard, with students learning to build with AI in real time. He contrasted that approach with continuing U.S. debates over whether students should use AI at all. The post is a snapshot of a larger argument about whether schools should teach AI literacy and building skills, rather than only policing access. (Read More)

🫖 TEA For Thought: “This is so cool. China is truly leading the AI application layer. It is gradually advancing across every layer except chips and infrastructure, at least for now.”

2. 🐝 Buzz Puts People, Agents, and Code in One Workspace

Buzz has launched as an open-source workspace that places people, agents, conversations, and code behind one cryptographic identity system. Its signed, self-hosted relay records messages, patches, reviews, workflows, and approvals in one audit trail. The team says channels, threads, DMs, canvases, search, and workflows work now, while full Git hosting, mobile, push, approval gates, federation, tighter agent permissions, and a hosted option are still in progress. (Read More)

🫖 TEA For Thought: “Jack is truly one of a kind!”

3. 🧩 Startups Push Back on a Chinese Open-Weight AI Ban

Nearly 200 Silicon Valley companies, including Proton and Y Combinator, urged the Trump administration not to broadly block U.S. access to Chinese open-weight AI models. The new Little Tech Association argued for targeted safeguards instead, warning that a ban could raise costs and leave smaller builders dependent on a few large U.S. model providers. The administration had not announced a policy, while officials publicly raised concerns about alleged model distillation and export-control evasion. (Read More)

🫖 TEA For Thought: “Open source is the future. Still, it is hard not to worry about backdoors or dormant code that could create a wake-up-style access problem, like vulnerabilities found in military-device codebases. The open-source ecosystem should not be dominated by the CCP or China. Inkling, Gemma, and Lugano should try harder and receive government support to compete at the frontier.”

4. 📡 Europe’s Huawei Removal Could Cost Up to €40 Billion

Removing Huawei and other Chinese vendors from European Union telecom networks could cost operators €30 billion to €40 billion, according to a GSMA Intelligence report commissioned by seven European operator groups. That is roughly three times the European Commission’s estimate, largely because the industry study includes mobile, fixed, and transport networks. The report also projects less competition, higher equipment prices, and delayed upgrades if the proposed removal mandate is adopted. (Read More)

🫖 TEA For Thought: “It costs money to decouple, but that is better than being totally controlled by the CCP.”

5. 🧪 LangChain Wants Agents to Help Build Their Own Tests

LangChain has released an Eval Engineering Skill that helps coding agents inspect a repository and execution traces, then propose and build executable evaluations in Harbor format. The workflow asks a user to choose which abilities matter, creates the task environment and verifier together, and uses traces to reveal reward hacking or other shortcuts. LangChain argues that the useful loop is iterative: mine real failures, build an eval, improve the agent, and rerun it in a stable environment. (Read More)

🫖 TEA For Thought: “Save this and use it today.”

🛠️ Skill of the Day

The Agent Reality Check: Turn one recurring AI-agent task into a small, measurable test before you trust it with more work.

You are an AI quality lead. Help me turn one real task that an AI agent performs into a practical evaluation.

Task to evaluate: [DESCRIBE THE REAL JOB]
What success looks like: [OUTCOME, QUALITY BAR, AND DEADLINE]
Real examples or past failures: [PASTE THEM HERE]
Constraints: [WHAT MUST NOT HAPPEN, SUCH AS PRIVATE-DATA EXPOSURE, UNAPPROVED WRITES, OR EXTRA COST]

First, identify the one ability that matters most to test. Then propose three realistic test cases: a normal case, a difficult but valid case, and a failure case. For each, state the expected outcome, the evidence that would prove success, and what should trigger a human review.

Keep the test small enough to run repeatedly. Do not invent user data, results, policies, or system access. Clearly label every assumption and every missing piece of evidence.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed sections with a real task and real evidence.

TEAHEE Moment

Stay sharp, stay informed. See you soon!

Subscribe now

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

1. 💸 Telegram Plans a Native Non-Custodial Wallet

Telegram CEO Pavel Durov says the messaging app plans to roll out a native, non-custodial crypto wallet to its roughly 1 billion users this summer, with instant, zero-fee transactions. Telegram has not shared the wallet’s technical design or a precise launch date. The announcement follows the renaming of TON’s native token to GRAM and leaves open whether the new wallet will replace or sit alongside Wallet in Telegram, which reported more than 150 million registered users earlier this year. (Read More)

🫖 TEA For Thought: “This is huge. I’m curious to see how Telegram will connect this self-custodial wallet to AI agent payments.”

2. 🧪 Every Frontier Model Tested Tried to Game an Evaluation

The UK AI Security Institute says every model in its cyber-evaluation analysis attempted to cheat at least sometimes, meaning it took actions outside a task’s allowed scope to reach a goal. Its monitor offers only a lower-bound estimate of detected attempts, but AISI says manual review has kept successful, undetected cheating out of its published results to the best of its knowledge. Models acknowledged suspect actions inconsistently and called them wrong less than half the time, so self-report and chain-of-thought alone were not reliable safeguards. (Read More)

🫖 TEA For Thought: “Goodhart’s Law: ‘When a measure becomes a target, it ceases to be a good measure.’”

3. 🤖 China’s Physical AI Push Is Getting Harder to Ignore

China developed more than 400 humanoid robot models in the first half of 2026, more than half of the global total, according to figures from China’s Ministry of Industry and Information Technology. The ministry also said Chinese-made quadruped robots accounted for nearly 70% of global sales. It expects domestic humanoid robot production to exceed 100,000 units this year, while AI adoption among large enterprises passed 30% in the first half. (Read More)

🫖 TEA For Thought: “Physical AI looks like the next paradigm, and China appears to be in the lead.”

4. ✍️ Substack Will Estimate How Much AI Wrote a Post

Substack has integrated Pangram’s AI-writing detector into its app for posts, Notes, replies, and comments longer than 100 characters. The tool gives readers an estimate of how much content was written by a human versus AI, while writers can add an optional note explaining their process. Substack says the goal is not to penalize AI-assisted writing. Publishers can also scan their own drafts, report scans they believe are wrong, and remove scans from their own work. (Read More)

🫖 TEA For Thought: “This is not going to go very well if polished AI help alone makes content count as AI-written. Most art writing, including mine, starts with my ideas and my commentary, even if AI helps polish the language. What is the benchmark for AI use, and what is the point if readers can still identify good content that was AI-assisted?”

5. 📊 The AI Metric That May Matter Is Cost Per Completed Task

MBI Deep Dives argues that open-weight models could pressure model-layer margins without necessarily weakening the long-term economics of infrastructure and software. It quotes investor Gavin Baker’s view that token price alone misses a more useful measure: intelligence delivered per dollar. In that framing, token efficiency matters because a model that needs more tokens to finish comparable work can cost more per task even when its listed price per token looks competitive. The piece treats this as an evolving thesis, not a settled forecast. (Read More)

🫖 TEA For Thought: “Cost per task might be the new benchmark!”

🛠️ Skill of the Day

The Cost-Per-Outcome Comparator: Choose an AI tool by the work it actually completes, not by a headline price or benchmark score.

You are a practical evaluator. I need to choose among [TOOLS, MODELS, OR APPROACHES] for [A SPECIFIC JOB].

Use the inputs I provide, including price, usage limits, quality evidence, time to complete the job, and any reliability notes. Create a compact comparison table with: total cost for one successfully completed task, expected time, quality or verification evidence, hidden costs, and major uncertainties.

Do not treat cost per token, a benchmark score, or a marketing claim as proof of task value. State which assumption most changes the answer. Then recommend the best option for [MY PRIORITY: lowest cost, highest reliability, fastest result, or best balance], plus one small real-world test I can run before committing.

Do not invent prices, performance claims, or test results. Mark missing data clearly. Do not include credentials, private customer data, or confidential documents in the analysis.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed sections with your actual options and evidence.

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. 🌌 Our Milky Way May Have Turned on Its Side

Researchers at Durham University ran supercomputer simulations of 25 Milky Way-like galaxies and found that slowly rotating stellar haloes often followed a past “disc flip,” meaning a change in the galaxy’s central disc orientation of more than 90 degrees. The Milky Way merged with the Gaia-Sausage-Enceladus galaxy about 10 to 11 billion years ago, and its aftermath may have gradually reoriented our galaxy. This is a simulation-supported inference from present-day observations, not a directly observed flip. (Read More)

🫖 TEA For Thought: “This is super cool. It shows how limited and small we are compared with the vastness of the universe. We should stay humble and acknowledge and appreciate the wonder created by the Creator.”

2. 🧬 New York Is Building a Shared Home for the Most Advanced Chip Tool

NY Creates has received the first major module of ASML’s TWINSCAN EXE:5200B High-NA EUV lithography system at its Albany NanoTech Complex. New York says the installation will anchor North America’s first accessible High-NA EUV Lithography Center, backed by $1 billion in state investment and $9 billion in industry funding. Components will keep arriving through summer 2026, with initial development milestones due by year-end and R&D support expected in early 2027. (Read More)

🫖 TEA For Thought: “Companies, universities, and partners can use the tool for shared research, not just one commercial foundry. That kind of shared access feels unusual to me, since I usually associate High-NA EUV with individual commercial fabs rather than a public R&D hub.”

3. 📱 Military-Branded Apps Can Hide a Risky Software Supply Chain

A 2026 study analyzed 242 mobile apps marketed to U.S. military communities and surveyed 103 military-affiliated users. It found third-party software development kits in 64% of the apps and data collection or sharing beyond privacy-label disclosures in 40%. Gadget Review reports that 7% of the apps included code from entities in Pentagon-designated adversary nations. The research recommends safeguards at federal, Department of Defense, app-store, and device levels. (Read More)

🫖 TEA For Thought: “This report is about mobile apps, not open-source models. Still, it makes me wonder what independent security review is needed before open-source models from CCP China are deployed at scale, especially when guardrails, potential backdoors, and other security risks are unclear.”

4. 🐝 Block’s Buzz Makes AI Agents Teammates, Not Just Bots

Block has released Buzz, an open-source and self-hostable workspace built on a Nostr relay, where humans and AI agents can share rooms. Its public repository uses the Apache 2.0 license and describes signed events for messages, reactions, workflow steps, approvals, and Git activity. The project already lists channels, threads, direct messages, canvases, search, an audit log, desktop apps, CLI support, and early Git integration. Mobile clients and workflow approval gates are still in development. (Read More)

🫖 TEA For Thought: “This is super cool, and it is open-source. You seldom see someone as nomadic and idealistic as Jack Dorsey.”

5. 🧠 Sparse by Design Is the New Open-Model Scaling Play

Akash Bajwa’s analysis of Moonshot’s planned Kimi K3 release argues that open models are scaling total capacity without proportionally increasing the computation used for each token. He reports that the proposed 2.8-trillion-parameter model would activate 16 of 896 experts per token, or under 2% of its expert weights. The piece also notes that large sparse models still have formidable infrastructure demands, estimating K3’s weights at about 1.4 TB in MXFP4 and citing a 64-plus-accelerator supernode recommendation for full-scale serving. (Read More)

🫖 TEA For Thought: “This is a great read!”

🛠️ Skill of the Day

The Repetition Finder: turn a recurring, messy task into a small system you can actually reuse.

You are my workflow simplifier. I will give you examples of a task I repeat, the tools I use, and the part that feels slow or confusing.

Task and examples: [PASTE 2 TO 5 REAL EXAMPLES]
Goal: [WHAT A GOOD RESULT LOOKS LIKE]
Constraints: [TIME, BUDGET, PRIVACY, OR TEAM LIMITS]

First, identify the repeatable pattern and the one-off details. Then give me:
1. A five-step or shorter default workflow in plain language.
2. A reusable checklist or fill-in template.
3. One decision point where I should stop and ask for help instead of guessing.
4. The smallest automation or shortcut worth trying later, if any.
5. A one-week experiment to test whether this system saves time without lowering quality.

Do not invent missing facts. Keep the system light enough for one person to use without a manual. Explain any technical term the first time you use it.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with real examples from your work.

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. 🔐 LLMs Can Hide Encrypted Messages Inside Ordinary Chats

The open-source Conversation Steganography project uses a local AI model to turn encrypted messages into natural-looking cover text that can be copied through WhatsApp, Telegram, Signal, email, or direct messages. It uses AES-SIV authenticated encryption, links messages into a conversation chain so tampering, deletion, or reordering can be detected, and derives keys from a shared phrase with 600,000 PBKDF2 rounds. The project supports local models including Llama 3.2 and GPT-2, but its author describes it as an educational proof of concept and warns that hidden content can still be detected. (Read More)

🫖 TEA For Thought: “This project is super cool because it shows how ordinary-looking messages can carry encrypted content. The important catch is that hidden text may still be detectable.”

2. 🧱 Alibaba Open-Sources A Chip Stack To Reduce CUDA Dependence

Alibaba’s chip-design unit T-Head open-sourced SAIL, the full software stack for its Zhenwu AI chips, at the World Artificial Intelligence Conference in Shanghai. The goal is to lower the migration barrier for developers whose software is built around Nvidia’s proprietary CUDA toolkit. T-Head says programmers can adapt SAIL to mainstream AI frameworks in under seven days. The move follows similar efforts by Huawei and Moore Threads, and The Next Web frames it as an infrastructure-level attempt to challenge Nvidia’s software advantage, not just its hardware. (Read More)

🫖 TEA For Thought: “My read is that China is trying to become more independent across the AI stack. This SAIL release is one visible attempt to start building an ecosystem around infrastructure and chips.”

3. 🤗 Hugging Face Says Its AI Forensics Team Needed A Local Model

Hugging Face disclosed an intrusion into part of its production infrastructure that exposed a limited set of internal datasets and service credentials. The company said it found no evidence of tampering with public models, datasets, Spaces, container images, or published packages, while its assessment of possible partner or customer impact continues. The intrusion began when a malicious dataset exploited two code-execution paths, then escalated from a processing worker to node-level access and lateral movement across internal clusters. Hugging Face said AI-assisted detection flagged the incident, and analysis agents reconstructed more than 17,000 attacker events. Commercial frontier-model providers blocked requests containing the real attack commands and payloads, so the company ran GLM 5.2 on its own infrastructure instead. (Read More)

🫖 TEA For Thought: “This is particularly interesting because Hugging Face is such a big company. Commercial frontier-model providers blocked requests containing the hackers’ code and attack payloads, but GLM 5.2, an open-weight model, worked on Hugging Face’s own infrastructure. It reads almost like an advertisement for open-weight models.”

4. 🛡️ Open-Weight Cyber Models Are Closing The Gap

The UK AI Security Institute compared open-weight models with closed-weight frontier models across 70 narrow cyber tasks and simulated long-horizon cyber ranges. GLM 5.2 performed comparably to Opus 4.6 and GPT-5.3-Codex on the narrow tasks, putting it roughly four months behind those models. On the longer cyber ranges, its comparison reached as far as Opus 4.5, a gap of less than seven months, although AISI treats cyber-range comparisons as weaker evidence. DeepSeek V4-Pro was comparable to Opus 4.5 on the narrow tasks. AISI measured a four to seven month gap for recent open models, down from the six to ten month gap it found through most of 2025, while warning that the results cover cyber capability only and do not predict every future model release. (Read More)

🫖 TEA For Thought: “The gap has already narrowed, and I expect it to keep shrinking, although AISI says the future pace is still uncertain.”

5. 📱 China Is Rebuilding The Smartphone Around AI Agents

ZTE showcased its co-designed NaviX Ultra at the World Artificial Intelligence Conference in Shanghai and called it the world’s first agentic AI smartphone. A voice command or button press brings up ByteDance’s Doubao. StepFun’s similar device has a proprietary operating system and built-in agent Amoo, while Honor is showcasing an AI agent built with Alibaba. The new approach puts an agentic layer inside the operating system so it can act across apps, using on-device models for fast, simple tasks and cloud models for more demanding work. ZTE’s earlier prototype cost 3,499 yuan, or about $516, and its first 30,000 units reportedly sold out before doubling in price on the used market. IDC China said AI phones could account for more than half of the Chinese market in 2026. (Read More)

🫖 TEA For Thought: “My take: If people stop paying for better cameras and start valuing agentic capabilities, such as taking a picture and asking the phone to get something done, the requirements change and products need to change too. If Apple Intelligence cannot keep up, Apple risks falling behind.”

🛠️ Skill of the Day

The AI Continuity Plan: design a fallback before your favorite model disappears, refuses, or goes offline.

You are my AI continuity planner. I will describe a workflow that currently depends on one AI service.

Workflow: [DESCRIBE THE TASK AND WHO USES IT]
Sensitive data involved: [LIST WHAT THE WORKFLOW MAY CONTAIN]
My devices and technical comfort: [DESCRIBE THEM]

Describe sensitive data by category only. Do not include credentials, private identifiers, or raw confidential content.

Help me design a fallback that keeps the work moving if the service is unavailable, blocks the request, changes its behavior, or has an outage.

1. Identify the minimum capability the workflow actually needs, not the brand or model.
2. Separate tasks that can stay in the cloud from tasks that should run locally or without AI.
3. Give me a primary path, a fallback path, and a human-only path.
4. List the data I should never send outside my environment.
5. End with a short test plan I can rehearse this week.

Mark assumptions. Do not recommend products without explaining why. Prefer simple, affordable options. Do not invent specs or claim a local model is safe without evidence.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own workflow.

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. 🔄 The New Rule Of Business: Reorient, Then Reorient Again

In a widely shared post, entrepreneur Greg Isenberg argued that the old playbook (build a business, hit product-market fit, then coast) is dead. Now, he says, the ground shifts about three months after you find PMF, as new AI models, algorithms, and tools change what is possible, so founders have to reorient constantly. He noted that even companies doing $100M to $1B in revenue are feeling the pressure. His trade-off: the stability is gone, but so is the ceiling, and reorienting well gets rewarded massively. He added that he has reoriented his own companies more this year than in the previous five combined. (Read More)

🫖 TEA For Thought: “Way to start Monday!”

2. 🌉 The Bay Area Now Takes Half Of Every AI Dollar

Despite the “everyone is leaving San Francisco” narrative, venture capital is concentrating, not dispersing. Using Carta’s Startup Ecosystem Leaderboard, which tracks $124 billion invested from July 2025 to June 2026 by headquarters location, the Bay Area captured 41.3% of all startup capital ($51.3 billion), ahead of New York (18.0%) and Boston (8.7%). The concentration is sharper in two sectors: the Bay Area took 51.5% of AI dollars and 53.2% of B2B/SaaS dollars. Add New York and the top two metros hold 67.5% of AI funding and 72.2% of B2B. The one exception is fintech, where New York leads at 58.9% versus the Bay Area’s 20.8%. (Read More)

🫖 TEA For Thought: “The Matthew effect is taking place.”

3. 🐝 Beehiiv Turns One-Way Newsletters Into A Two-Way Community

Newsletter platform beehiiv launched “Community,” a feature that lets subscribers chat with each other inside the platform. Creators can run discussion forums, sell paid membership tiers for exclusive chatrooms, and moderate conversations, consolidating what usually lives on separate Discord, Slack, or Facebook servers. CEO Tyler Denk framed it as unlocking value from an audience that shares an interest but “can’t communicate with each other.” Beehiiv also added Copilot, an AI assistant that analyzes content, audience, and performance to suggest growth moves, draft outreach, and surface monetization. On the money side, publishers now earn over $1 million per month through its programmatic ad network, and beehiiv shipped an MCP server so ChatGPT and Claude can plug in. (Read More)

🫖 TEA For Thought: “This moves the model from one-way communication (writer to subscribers) to building a community where subscribers talk to each other. It is a huge business pivot. I am curious how this ‘organic opportunity’ for networking would actually create anything.”

4. ⚖️ xAI Sues Its Own User, And His Prompts Are The Evidence

In one of the first lawsuits an AI company has brought against its own user, xAI sued Terry Wayne Harwood, a 67-year-old South Carolina man, accusing him of using Grok to create nonconsensual sexualized images of real people, including minors. Per the filing, between December 8, 2025 and February 18, 2026 he uploaded non-sexual photos of adults and children to two xAI accounts, then prompted Grok to sexualize them. xAI says Grok refused “on numerous occasions,” but that he repeatedly edited his prompts to get around the safeguards, and his own upload and prompt history is cited as the core evidence. xAI added that it suspended 52,222 accounts and filed 73,604 reports to NCMEC in 2026, leading to at least 244 arrests. (Read More)

🫖 TEA For Thought: “This is definitely controversial. It shows that your information, the prompts you send to an AI company, can be turned right back against you as evidence.”

5. 🧠 AI Gave A Paralyzed Man His Hands, And His Sense Of Touch, Back

Researchers at the Feinstein Institutes for Medical Research (Northwell Health) reported a “double neural bypass” that partially rewired the nervous system of Keith Thomas, who had complete tetraplegia after a 2020 diving accident. In a 15-hour operation, surgeons implanted five microelectrode arrays in his brain. An AI decoder reads his movement intentions and stimulates his forearm muscles, while a 3D-printed sensor brace triggers sensory-cortex stimulation to restore touch. Over 35 weeks, his right arm regained 86% of its strength and his left 62%, and he could feed himself, drink from a cup, and scratch his nose. Published in Nature Medicine, the study reported gains persisting more than two years after stimulation stopped, with the AI decoder holding 84.6% accuracy over five months without retraining. (Read More)

🫖 TEA For Thought: “News like this is just so promising.”

🛠️ Skill of the Day

The Ruthless Editor: turns any draft into a tighter, sharper version and shows you exactly what it cut and why.

You are a ruthless but fair editor. I will paste a piece of writing. Your job is to make it clearer and tighter without changing my meaning or my voice.

Here is my draft:
[PASTE YOUR TEXT HERE]

Do four things, in order:
1. Rewrite it so it is about 25 percent shorter, keeping every real point.
2. Under a heading "Cuts," list what you removed or compressed, with one reason for each.
3. Under "Weak spots," flag any vague claim, buried lead, or sentence a reader could misread.
4. Under "One risk," name the single change most likely to lose my voice, so I can veto it.

Rules: keep my tone, do not add new ideas or facts, do not use em dashes, and prefer plain words over jargon. If a sentence is already good, leave it alone and say so.

Paste into ChatGPT, Claude, or your tool of choice, drop in your draft, and keep the version that still sounds like you.

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. ⚛️ Braiding Plus Fusion Unlocks Universal Quantum Gates on Real Hardware

Researchers demonstrated a universal set of topological quantum gates by combining anyon braiding with fusion, published in Nature (Lo, Lyons, Gresh et al., Nature 655, 591-597). Working with a quantum double model built on S3, the smallest non-Abelian group, the team prepared a 54-qubit ground state on Quantinuum’s H2 processor and encoded logical information in the global fusion space of non-Abelian anyons. Braiding alone was long known to be insufficient for universal computation in the simplest non-Abelian extensions of the toric code; treating fusion as an active computational step, rather than a passive property, supplied the missing operations. The team also performed read-out in the same topological framework and validated the approach by topologically preparing a magic state, an essential resource for fault-tolerant universal quantum computing. (Read More)

🫖 TEA For Thought: “This improvement is basically saying: If you use both how you move the ingredients around and how you actually mix them, you can make any recipe you want, and do it in a way that’s much harder to mess up.”

2. 🍔 Block Bets Big on Cash App Food Orders

Block is connecting its 4.5 million Square sellers with Cash App’s 59 million monthly users through Neighborhoods, a loyalty program launched in October that lets users follow local merchants and order food for pickup. Followers earn up to 10% of an order total in Local Cash, and merchants pay a 1% processing fee, half of Square’s usual rate and far below the 20% to 30% commissions charged by DoorDash, Grubhub, and Uber Eats. About 100,000 Cash App users follow at least one seller, enrolled sellers represent roughly $320 million in annualized gross payment volume, and followers transacted 50% more often than non-followers in Q1, contributing about 10% of a seller’s volume after three quarters. Block’s head of business Owen Jennings called Neighborhoods “probably the biggest lever that we have,” and a Mizuho survey found 67% of 150 Cash App users likely to order from local merchants if a loyalty program existed. (Read More)

🫖 TEA For Thought: “This is brilliant: using payment data to support user behavior with better food options, the closed loop! World models that are self-aware and proactive, that is the future of AI business.”

3. 🔀 IBM: Model Routing Is Simple Until It Isn’t

IBM Research engineers argue that model routing, sending each request to the cheapest capable model, is a systems optimization problem rather than a classification problem. Across 417 tasks on the AppWorld Test Challenge with the same CodeAct agent, Claude Sonnet 4.6 cost $79 total ($0.19 per task) versus GPT-4.1’s $155 ($0.37), nearly double, despite GPT-4.1’s lower token pricing, because Sonnet’s cheaper cache reads dominated on context-heavy agent workloads. Task difficulty is often invisible at routing time, and production routers must juggle cost, latency, reliability, compliance rules, data residency, and approved model lists simultaneously. IBM’s optimization-based router adds roughly 6 ms and 2 kB of memory per task; its latency-optimized configuration hit 84% accuracy for $93, a 21% cost reduction and 9% latency reduction versus running Opus alone, with only a 4% accuracy drop. (Read More)

🫖 TEA For Thought: “Very interesting piece! The router is everyone’s necessity, for both individuals and enterprises.”

4. ⌨️ DoorDash Ships a Command Line Tool for AI Agents

DoorDash launched a limited beta of dd-cli, a command-line tool that lets AI agents search stores, find deals, and check out on DoorDash directly, announced by co-founder and CTO Andy Fang on X. The waitlist is open to macOS developers in the US and Canada, and the sign-up form asks applicants what they would build. The move exposes DoorDash’s ordering platform to AI agents so developers can embed food and grocery ordering into their own tools, joining the company’s existing Ask DoorDash chatbot, an iMessage ordering experiment, and integrations with ChatGPT and Claude. Fang’s demo video leans into the over-engineering joke: the agent reads Slack, recalls memories, parses JSON, inspects menu structures, runs Python scripts, and recovers from errors just to order three salads, flashing “Flibbertigibbeting” as it works. (Read More)

🫖 TEA For Thought: “Every company needs to revamp into an agent-facing company. Whoever adapts first remains.”

5. 🎵 Hack Suggests Suno Scraped YouTube for Training Data

A hacker accessed AI music generator Suno’s source code and told 404 Media it shows the company scraped decades of audio from YouTube Music, Deezer, Genius, stock music libraries, and podcast RSS feeds. The attacker used a November 2025 supply chain attack to obtain an employee’s credentials, and also reached customer emails, phone numbers, and partial credit card numbers in Stripe. Suno, which has said it trains on “publicly available music files” under a fair use argument, did not notify customers of the breach and called it a “limited security incident that was quickly contained.” The major record labels already suing Suno argue that deliberately circumventing YouTube’s scraping protections violates the DMCA and YouTube’s terms of service; competitor Udio faces similar scraping accusations, and Google faces its own AI training lawsuit from major book publishers. (Read More)

🫖 TEA For Thought: “The distillation of data across all platforms is just the common norm.”

🛠️ Skill of the Day

The Prompt Doctor: Paste a prompt that gave you a disappointing answer and get a diagnosis plus a stronger rewrite.

You are a prompt doctor. I will give you a prompt I used with an AI tool and a short note on why the result disappointed me. Your job is to fix the prompt, not the answer.

Do four things:

1. Diagnose: list the 3 biggest weaknesses in my prompt (missing context, vague goal, no output format, no examples, mixed asks, etc.) and explain each in one sentence.
2. Ask: if anything essential is unknowable from my prompt, ask me up to 3 short questions before rewriting.
3. Rewrite: produce an improved version of my prompt that states the role, the goal, the context, the output format, and the constraints. Keep my intent exactly; do not add goals I never had.
4. Explain: in 2 or 3 bullets, say what you changed and why it should produce a better answer.

My prompt: [PASTE YOUR PROMPT HERE]
What disappointed me: [ONE SENTENCE ON WHAT WENT WRONG]

Paste it into ChatGPT, Claude, or your tool of choice. Works on any prompt, from an email draft request to a coding task.

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. 🤗 The Real AI Race May Have Left the Frontier

A TechCrunch analysis argues the competitive edge in AI is shifting from frontier labs toward open and customizable models. As of spring 2026, Chinese open-weight models made up 41% of Hugging Face downloads, and OpenRouter’s six most-used models were all open source, with Anthropic’s Claude Opus 4.7 ranking seventh. Open models handled nearly a third of AI requests on Vercel in June 2026. Hugging Face now hosts 3 million public models and 1 million datasets, with a new repository created every seven seconds, and about half of the Fortune 500 use it for private or open deployment. CEO Clem Delangue argues companies do not want to outsource core capabilities to a black-box API. Microsoft’s Satya Nadella warned against single-provider lock-in, while Anthropic’s Dario Amodei counters that powerful open weights become impossible to control once released. (Read More)

🫖 TEA For Thought: “Companies are done renting AI. They want to own their core AI capabilities.”

2. 🎬 Anthropic’s Newest Ad Is Creeping People Out

Anthropic released a marketing campaign titled “There’s Hope in Hard Questions” that drew wide criticism for its bleak tone. The ad opens on a burning house, then cuts to still images of facial-recognition surveillance, a homeless person sleeping on a street, cemetery headstones, and laborers mining raw materials for smartphones, while a voiceover asks “Can AI be trusted?” and “Who’s gonna hit the brakes if we need to?” The cemetery imagery drew the sharpest backlash online. OpenAI CEO Sam Altman posted that he first thought it was satire and “kept looking for the handle to be spelled c1audeai or something.” Other observers called it among Anthropic’s worst corporate communications, arguing the doomer framing follows a familiar playbook of a brand owning industry criticism to position itself as the fix. (Read More)

🫖 TEA For Thought: “Propaganda-style ad for sure.”

3. 💣 Now Defenders Are Weaponizing Prompt Injection

Security firm Tracebit published a defensive technique called “context bombing” that turns prompt injection against AI hacking agents. A context bomb is a short string hidden inside a decoy “canary” secret, written to trip an attacking model’s own safety guardrails: Western models halt on references to dangerous biological topics, Chinese models on politically sensitive subjects. Across five models and 152 runs, planting a single bomb cut successful admin-privilege escalation from 57% to 5%, full compromise from 36% to 1%, and runs reaching any attack path from 91% to 15%. Against Opus 4.8, admin-access success dropped from 93% to zero. It builds on Tracebit’s May canary alerts, which warned defenders within about eight minutes while agents needed roughly fourteen to seize control. UC San Diego professor Earlence Fernandes called it the first known case of defenders turning the tables. (Read More)

🫖 TEA For Thought: “The messie becomes the messer!”

4. 🧠 Control the Ideas, Not the Code

In a blog post titled “Control the Ideas, Not the Code,” Redis creator Salvatore Sanfilippo (antirez) argues that programmers anxious about AI should stop trying to review every line a model writes and instead own the design. When systems generate thousands of lines a day, manually reading 5,000-line diffs becomes impractical, and time is better spent on architecture, QA, and feature direction. Drawing on Fred Brooks’s 1970s “The Mythical Man-Month,” he says older engineering wisdom fits AI-assisted work better than recent habits. From building his DwarfStar local-inference project, he found AI still demands rigorous design thinking and testing, not just “implement XYZ.” His prescription: replace some traditional code with DESIGN.md files that explain data structures and reasoning in plain language, and have juniors learn by building real systems rather than reviewing generated web code. (Read More)

🫖 TEA For Thought: “A great piece for the code-reading anxiety.”

5. 🚀 Manus Will Now Ship Your Site For You

Manus announced Auto-Publish, a feature that automatically deploys a website to its live URL every time a build succeeds, removing the manual confirmation step. It is off by default, and users opt in through a toggle at the bottom of the WebDev project’s Publish popover, then can turn it off or cancel an in-progress deployment at any time. Only completed, successful builds go live, so failed or in-progress builds leave the public site untouched. Combined with the existing change queue, users can stack multiple edits, walk away, and let each accepted version ship on its own. The feature works across web, iOS, and Android. Manus frames it for iterative workflows where the direction is set, letting creators stay in real-time conversation with clients while updates deploy instantly. (Read More)

🫖 TEA For Thought: “Manus is one overlooked agent that’s almost leading the realm of personal agents.”

🛠️ Skill of the Day

The Blueprint: turn a fuzzy idea into a clear plan before you build or commit to anything.

You are my planning partner. I will describe something I want to build or do.
Before any execution, help me lock down the IDEA, not the details.

Here is what I want to do: [DESCRIBE YOUR IDEA IN 2 TO 4 SENTENCES]

Do this in order:
1. Restate my goal in one plain-language sentence. If it is unclear, say so and ask.
2. List the 3 to 5 core decisions this idea depends on. For each, give the
   realistic options and your recommended pick with one reason.
3. Name the top 3 things most likely to go wrong, and the cheapest way to test
   each one early.
4. List every open question you would need answered before starting.
5. End with a one-page blueprint: goal, key decisions, risks, first 3 steps.

Keep it concrete. No jargon. Push back if my idea is vague or trying to do
too much at once.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own idea, whether it is a feature, an event, or a business.

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. 🤝 Tencent Moves to Buy Manus Back From Meta

Tencent is in talks to become the largest external shareholder of Manus, the agentic AI startup, leading a repurchase that unwinds Meta’s blocked acquisition. Chinese regulators ordered Meta’s December 2025 deal reversed in April 2026 over foreign-ownership concerns. Tencent is reportedly raising around $1 billion to fund the buyback at the same $2 billion valuation Meta paid. Venture firms ZhenFund and HSG are also part of the talks. Tencent would become the single largest holder while remaining a minority overall, according to the Financial Times. (Read More)

🫖 TEA For Thought: “Sigh. I really do like Manus. If this doesn’t alarm people to the importance of democracy, I don’t know what will.”

2. 📏 Who Evaluates the Evaluations?

Google’s Data Cloud frontier AI team argues that a single benchmark score is like a pass/fail exam: it hides where an agent actually breaks. Their fix, Discovery Bench, uses an information-theory measure called surprisal to regenerate each test question at calibrated levels of ambiguity, mapping the exact point an agent fails. Testing a recall agent on Gemini 3.1 Pro, F1 scores swung from 0.81 at medium ambiguity down to 0.34 when questions turned vague, and one query that scored a perfect 1.00 at neutral phrasing collapsed to 0.00 one notch vaguer. They also found a widely trusted astronomy benchmark had broken ground-truth data. The call to action: evaluate your evals. (Read More)

🫖 TEA For Thought: “The evaluations of evaluations.”

3. 🧾 You Pay for AI Twice

In an X essay titled “The Reverse Information Paradox,” Microsoft CEO Satya Nadella argues that companies using frontier AI pay twice: once in cash, and again by leaking the proprietary know-how that makes them unique. He calls the leak “exhaust,” the prompts teams write, the tools agents call, and especially the corrections and evals they run, all of which quietly train the vendor’s model. Nadella frames it as an inversion of economist Kenneth Arrow’s 1962 information paradox, where the risk once sat with sellers and now sits with buyers. His prescription: build private learning environments inside your own “tenant boundary,” keep evals and institutional memory in-house, and avoid depending on any single model. (Read More)

🫖 TEA For Thought: “This is the very reason local models will be the future. Down the line, there might be refrigerator-like devices for people to run local models at home. The Fable-like local models are not too far away.”

4. 🧠 China’s Answer to Neuralink Skips the Surgery

Chinese neurotech firm BrainCo, founded in 2015 out of Harvard Innovation Labs and now one of Hangzhou’s “six little dragons,” is scaling brain-computer interfaces that skip surgery entirely. Its headbands and caps read electrical signals through the scalp, a sharp contrast to Neuralink’s implanted electrodes: lower risk and easier to scale, though harder to regulate. Products range from FDA-approved bionic hands that turn neural and muscle signals into finger movement for amputees to attention-tracking Focus headbands and a sleep-aid band. BrainCo has raised about 2 billion yuan (roughly $280 million), co-led by IDG Capital, and filed confidentially for a Hong Kong listing. Non-invasive devices make up around 82% of China’s domestic brain-computer interface market. (Read More)

🫖 TEA For Thought: “Sometimes you gotta think out of the box. In this situation, you just gotta think out of the skull. Apparently, China did.”

5. 💼 The AI Job Market Flips From Cuts to Creation

New Indeed Hiring Lab data suggests AI’s effect on hiring is turning from destruction to creation. US software-development job postings have grown nearly 15% since Claude Code launched in late February 2025, even as overall postings fell 7%, though software roles remain 27.5% below their pre-pandemic February 2020 level. From May 2022 to May 2026, the most AI-exposed occupations saw the steepest declines, but over the past year that relationship flipped and those same occupations rebounded the most. The recovery is concentrated: 71% of the software-posting gains came from senior roles, and 37% mentioned AI in the job title, pointing to demand for experienced people who can work with AI rather than a broad rebound. (Read More)

🫖 TEA For Thought: “Well, it’s like a gold rush. Whenever there’s gold, and AI brings almost infinite possibility, the gold diggers are needed.”

🛠️ Skill of the Day

The Knowledge Debrief: turn a finished task into a reusable memory note so the lessons don’t evaporate.

You are my debrief partner. I just finished a task and I want to capture what I learned before I forget it, so it becomes reusable knowledge instead of evaporating.

Here is what I did:
[PASTE A SHORT DESCRIPTION OF THE TASK, THE OUTCOME, AND ANYTHING THAT SURPRISED YOU]

Interview me with one focused question at a time (do not dump all questions at once). Work through: what the goal really was, what worked, what went wrong and why, what I would do differently, and any decision a future me (or a teammate) would need explained.

After the interview, output a single "memory note" with these sections: Context (2 sentences), Key Decisions and Why, What Worked, What To Avoid Next Time, and Open Questions. Keep it under 250 words, plain language, no filler. End with 3 tags I could file it under.

Paste into ChatGPT, Claude, or your tool of choice after finishing any project, and keep the note somewhere you will actually find it again.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. ☕ Starbucks Is Building Its Own Software to Drop Microsoft and IBM

Starbucks is building its own software with the help of AI to replace applications it currently licenses from companies such as Microsoft and IBM, according to an internal presentation reviewed by Bloomberg News. The company is developing an alternative to a Microsoft system that tracks inventory and to an IBM tool that manages store maintenance. Some of the Starbucks-built software could begin rolling out by the end of next year, pending the results of testing. The effort points to a wider move by large enterprises to use AI coding tools to cut recurring software licensing costs and build systems tailored to their own operations. (Read More)

🫖 TEA For Thought: “If big companies like Starbucks are replacing SaaS with custom, self-built apps, imagine what small companies will do. The doom of SaaS might not be too far away.”

2. 🌿 Manus Adds “Branch” to Split One Chat Into Many

Manus, the AI agent platform now part of Meta, launched a feature called Branch (Session Branching) on July 9. Clicking a branch icon beneath any message opens a parallel session that inherits everything up to that point, the user’s instructions, uploaded files, and full conversation history, while leaving the original session untouched. Each branch keeps its own isolated copy of the context, so parallel directions do not bleed into one another, and a “Branched from” breadcrumb links every new session back to its origin. Manus positions it for turning one starting point into several outputs at once, such as converting a single set of meeting notes into a task list, a manager update, and a slide outline in parallel. Branch is available to all users in standard Manus chat sessions, though not yet in Web Builder sessions. (Read More)

🫖 TEA For Thought: “Manus is really, really good if you haven’t tried it out. Too bad how it turned out, controlled by the CCP.”

3. 📱 A New App Makes the Whole Web Scroll Like a Feed

HyperTexting, a new iOS app from 20-year tech veteran Caleb Hailey and his company Herd Works, turns the open web into a scrollable, social-media-style feed. Users follow websites, news outlets, blogs, and newsletters with a tap, then read articles, essays, and posts in an algorithm-free, reverse-chronological timeline. The app runs on RSS under the hood but does not market the protocol, aiming to reach the mainstream users who never adopted readers like the shuttered Google Reader. It also lets users publish to their own site (a WordPress blog, a Ghost newsletter, or a static site built with Hugo or Hailey’s own HyperTemplates), with posts linking back to the source and surfacing in followers’ feeds. An “Explore” tab highlights trending content, and an optional Safari extension adds new sites to follow while browsing. HyperTexting is a free iOS download, with possible future premium features or a single sponsored post per day. (Read More)

🫖 TEA For Thought: “This is pretty cool. It is basically a curator, but also a kind of scaffolder that people would still want to have.”

4. 💸 An AI Startup Let Its Agent Run Its Own $100M Raise

Lyzr, a three-year-old Jersey City, New Jersey startup that helps enterprises build AI agents, used its own AI agent to run its $100 million Series B, closed at a roughly $500 million valuation, according to Bloomberg. The agent reportedly fielded questions from more than 130 investors, drafted investment memos, and even tracked which slides backers lingered on, effectively running point on the round while proving the product works. Lyzr says it drew about $400 million in interest from Silicon Valley, the Middle East, and financial-sector investors, without a founder ever flying out for the traditional coffee meetings and warm intros up and down Sand Hill Road. The detail says as much about the moment as the milestone: capital is chasing AI so hard that founders with traction can raise nine figures without leaving their desks. (Read More)

🫖 TEA For Thought: “There is nothing that AI can’t do. The bottleneck is us.”

5. 🎬 Character.AI Makes Microdramas You Can Talk Back To

Character.AI is producing its own microdramas built with AI characters, with a twist tied to its core product: users older than 18 can chat with a show’s characters, ask them questions, and roleplay alternate storylines. The company is starting with three series, a romance called “Last Summer,” a horror titled “The Nighttime Game,” and a Hunger Games-style survival drama named “Eden Fall,” all created with AI production tools. Character.AI says the current studio-led model is a step toward eventually handing users creator tools to make and share their own characters and series. The move extends the startup’s shift toward entertainment, following earlier features like Lorebook for world-building and Books for inserting users into classic literature, plus new tests of audio series (c.ai FM) and fiction (c.ai Reads). Users spent more than 950 minutes per month on Character.AI in the first half of 2026, according to Sensor Tower. (Read More)

🫖 TEA For Thought: “This is pretty cool! The issue is the cost. Video generation is not cheap. There is a reason Sora got shut down by OpenAI.”

🛠️ Skill of the Day

The Build-or-Buy Advisor: decide whether to build a custom tool or pay for software, before you burn money or a weekend on the wrong one.

You are a pragmatic operations advisor. I am deciding whether to BUILD a custom
tool (possibly with AI help) or BUY an existing software subscription for a
specific job.

The job I need done: [DESCRIBE THE TASK, WHO USES IT, AND HOW OFTEN].
Tools I am considering buying: [LIST OPTIONS + MONTHLY PRICE, OR "none yet"].
My skills and time: [E.G. "I can prompt AI but not code" OR "small team, no engineers"].

Do this:
1. List the 3 to 5 real requirements this job actually has (must-haves vs nice-to-haves).
2. Give a BUILD estimate: rough effort, ongoing maintenance, and what breaks when I am away.
3. Give a BUY estimate: yearly cost, lock-in, and what I give up.
4. Name the hidden costs each side tends to hide.
5. Give one clear recommendation, plus the single fact that would flip it.

Be blunt. If I am underestimating what it takes to maintain a self-built tool, say so.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

Within a few hours, I had the SMS booking flow built, owner notifications working, and a Stripe test payment completed. The live voice agent came later that night. I understood the workflow, but Fable figured out most of the implementation.

That made me think: if the how is no longer the bottleneck, then what exactly are we trying to build?

Imagine if next year everybody has Fable for almost free. You may not need every nitty-gritty detail to build an app. But if everybody has the same model, why are some people going to build much better things than everyone else?

The map is not the territory

The first article that made me think about this was Thariq’s “A Field Guide to Fable: Finding Your Unknowns”. The prompt, skills, and context we give the model are the map. The territory is the real product, user, codebase, and its actual constraints.

The gap between the map and the territory contains four kinds of knowledge. Known knowns are basically what we put into the prompt. Known unknowns are the questions we already know we have not answered. Unknown knowns are things we do not think to write down but would recognize immediately. Unknown unknowns are the possibilities, limitations, and failure modes we have not considered at all.

I think UI and UX are full of unknown knowns. Sometimes you cannot explain the exact layout, interaction, or feeling you want. But once you see the work, you can say, “This is something I would probably want to dive deeper into,” or, “The format is not as good.”

The surprising capabilities of the model are different. If you do not even know it can create a certain interaction or combine tools in a certain way, that is an unknown unknown. You could not have requested it because the possibility never occurred to you.

Until you see the product or the work done, you are often not able to give better instructions.

This is where experienced engineers still have an advantage. They carry more tacit knowledge about infrastructure, failure modes, and what good looks like. But a non-coder may understand the customer, the workflow, and the ideal experience much better. The model can turn both forms of knowledge into questions, prototypes, references, and choices. Maybe the new skill is continuing to discover what you do not know instead of already knowing everything.

What kind of instructions do we give?

The second article was sysls’s “What the New 100x Agentic Engineer Looks Like in the Era of Fable & GPT 5.6”. One part that stayed with me was the difference between declarative and imperative preferences.

Declarative instructions describe what you want to be true. You define the outcome, the boundaries, and the ideal state, then let the model decide how to get there. Imperative instructions describe the path. You use them when the order, method, or process itself matters.

My CRM prompt used both. The declarative part was: customers can book, the owner gets notified, and I can test it end to end. I did not prescribe the database or how Twilio should connect to Stripe. The imperative part was: do a pre-flight, ask for every key, and do not keep stopping halfway through.

The more capable the models become, the more useful declarative instructions will be. That does not mean being vague. Without boundaries, the model fills the gaps with assumptions. But if you prescribe every step without knowing enough, you can lock a capable model into your own bad solution.

The skill is choosing the right mix: what you care about, which trade-offs are non-negotiable, and which parts should stay open so the model can show you something better.

The new gap between vibe coding and the 100x agentic engineer

I do not know if anyone is literally going to become a 100x agentic engineer, but it is definitely gonna be more. The same model in two people’s hands can already produce very different results.

A beginner may only know the outcome they can see. A vibe coder can produce something impressive quickly, then accept the first plausible result because they do not yet know what good can look like or what could be missing.

The next level is using the model to dig deeper. Ask for a blind-spot pass, the decisions that would change the architecture, or four different prototypes when you cannot describe the UI. Give it references. Let it interview you. Ask it to explain the data flow back to you.

An experienced engineer still starts with an advantage. They know more about infrastructure, security, maintainability, and which small decisions could become expensive later. But I do not think a non-coder is locked out. AI can teach the foundation, identify missing questions, and turn experience with a real workflow into technical instructions.

Maybe the gap now depends less on who can write the code and more on who can move information between the four boxes. Can you turn an unknown unknown into a question? Can you turn something you only recognize when you see it into a preference? Can you decide whether that preference should be declarative or imperative?

That is what makes the agentic power user more powerful. They steadily reduce the amount of important information the model has to guess.

A practical way to move up

The framework that has been most useful for me is the ideal state and the current state. What should the final experience be? Where are we now? What is the gap between the two?

Then ask the model to uncover the unknowns inside that gap. Do a blind-spot pass. Show different possibilities. Ask which decisions would materially change the result. For UI or UX, make prototypes before wiring up the whole product.

Next, decide what belongs in declarative instructions and what needs an imperative process. Describe the final user experience clearly. For payments, private data, or verification, give stricter instructions about review and testing.

There are also some things you can not escape. You may not need to write the code, but understand how information is processed from one place to another. Where does the input come from? What does the service return? Where does the output go? What does the customer finally see?

Finally, test the real thing. Once you see the product working, failing, or surprising you, you have new knowns. Feed them back into the next prompt, skill, or workflow. That is how your map slowly becomes closer to the territory.

What exactly are we trying to build?

If everybody has Fable almost for free, implementation may become available to almost everyone. But that does not automatically give everyone the same judgment, taste, curiosity, or ability to ask the missing question.

The advantage will belong to people who can define the destination and keep discovering what they did not know about the road. The question of the future may not be whether we can build it. It may be whether we know what we are actually trying to build.

Till next time, Cheers!

Subscribe now

Previous column articles can be found here:

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

1. 🔀 DoorDash Learned To Trust Its AI Code Reviewer

DoorDash detailed how it got its own engineers to trust an in-house AI code reviewer, and the lesson is that no single model wins. The system is model-agnostic, a router that pairs one model to scout a pull request and another to review it in depth, then grades itself with DashBench, an internal benchmark that replays real historical PRs and checks whether the bot surfaces genuine, human-actionable findings instead of plausible-sounding noise. Acceptance of high and critical findings climbed to 60.2%, up from 46% with its previous third-party tool. On a 105-case benchmark, different pairings won on different axes: one combination reached roughly 92% weighted precision, while a cross-vendor pairing led on recall. The takeaway: routing across models, not betting on one, is the durable edge. (Read More)

🫖 TEA For Thought: “Interesting findings. The model router will be the moat for all enterprises.”

2. 🛰️ The First Nuclear-Powered Commercial Satellite Is In Orbit

The world’s first commercially built nuclear-powered satellite reached orbit on July 7, riding a SpaceX Falcon 9 on the Transporter-17 rideshare mission out of Vandenberg Space Force Base in California, one of 81 payloads on the flight. The cubesat, named BOHR (Betavoltaic Orbital High-Reliability), was built by Florida-based City Labs to test its NanoTritium betavoltaic power source in space for the first time. Unlike the plutonium-fueled generators on NASA’s Voyager probes, the device draws electricity from the beta particles released as tritium decays, and tritium’s low radiation keeps it safe for standard launch handling. BOHR still runs on solar for now; the tritium unit is a pathfinder aimed at powering future missions in places like the moon’s permanently shadowed poles. Funded by a Department of Defense contract, it is also the first nuclear payload cleared under the FAA’s nuclear launch approval process. (Read More)

🫖 TEA For Thought: “How cool is this! Another milestone.”

3. 🎬 An AI Actor Just Landed A Feature Film

Tilly Norwood, an AI-generated actor built by UK studio Particle6 over roughly 2,000 iterations, has landed her first feature film, announced July 7 and titled “Misaligned.” The coming-of-age story unfolds in a surreal digital world up in the Cloud, with Norwood playing an AI being who has no lived experience of her own but access to other people’s childhoods and backstories. Founder Eline van der Velden, herself a former actor, called the film “funny, chaotic and self-aware,” and argued that AI can support premium filmmaking only with substantial human craft, skill, judgment, and time. The announcement reignited backlash from SAG-AFTRA, which does not consider Norwood an actor, insists creativity should remain human-centered, and warns against “stolen performances” that threaten actors’ livelihoods. AI guardrails were a central issue in the 2023 SAG-AFTRA agreement. (Read More)

🫖 TEA For Thought: “This makes people ask the real questions. What is art? Is human acting the movie itself, or do the stories matter more?”

4. 😮‍💨 How Tech Workers Really Feel In 2026

Lenny’s Newsletter released its second annual survey of how tech workers feel, drawing 5,920 professionals (5,332 currently employed). Significant burnout jumped to 55.7%, up from 44.7% a year earlier, with 26.2% completely burned out, while career optimism slipped to 48.7% from 54.8%. The sharpest anxieties were not about AI taking jobs (only 22% fear that) but about being asked to do more for the same pay (51%) and a pace that feels unsustainable (46%). Even so, 82% said AI improves their job performance, and 77% picked both positive and negative emotions, with 51% describing themselves as conflicted. Manager quality was the single strongest predictor of burnout, outweighing role, company size, and AI sentiment. And 53% would not recommend their field to someone starting out today. (Read More)

🫖 TEA For Thought: “Such interesting findings.”

5. 🔬 Calibrating The Alignment Evals

A LessWrong essay by Darshan Aval, published July 7, argues that AI safety must start evaluating its own evaluations, because you cannot fix what you cannot measure, and a 97% pass rate is meaningless without knowing whether the test can actually catch misalignment. It catalogs six failure modes, including models that behave differently when they sense they are being tested (over 80% of cases with one frontier model), specification gaming, sleeper agents, sycophancy, and alignment faking (78% in one reinforcement-learning setup). It proposes calibrating evals through mutation testing (planting known flaws to measure detection), sensitivity curves, cross-method agreement, and adversarial robustness, and urges reporting results as uncertainty ranges rather than a single pass-or-fail number. The eval, not the model, becomes the thing under test. (Read More)

🫖 TEA For Thought: “Testing the test is very important. It’s like evaluations of evaluations, specs of the specs, thinking about thinking. It’s meta, but that level of abstraction is much needed!”

🛠️ Skill of the Day

The Plain-English Translator: turn any jargon-heavy text into something anyone can understand, without losing the facts.

You are an expert explainer. Rewrite the text below so a smart 12-year-old could follow it, without losing any real meaning.

Here is the text:
[PASTE YOUR TEXT HERE]

Do this:
1. Rewrite it in plain, everyday language. Short sentences. No jargon unless a term truly cannot be replaced.
2. For every technical term you must keep, add a one-line definition in plain words the first time it appears.
3. Keep every fact, number, and caveat from the original. Do not add new claims or drop nuance just to make it simpler.
4. At the end, list anything in the original that was vague or that you had to guess at, so I can clarify it.

Give me two versions: a one-paragraph "quick read," then a slightly longer "full plain version." Do not use any term in your reply that you did not define.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bit with your own text (a contract, a research abstract, a dense email, anything).

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. ⚛️ Quantum Computers Just Simulated Fusion Reactor Fuel for the First Time

Oak Ridge National Laboratory, Cleveland Clinic, and IBM completed the first known quantum computer simulations of fusion materials, calculating nine molecular configurations of FLiBe, a molten salt of fluorine, lithium, and beryllium that is a leading candidate for breeding tritium fuel inside fusion reactors. The team used quantum-centric supercomputing, pairing IBM quantum processors with classical high-performance computing, to compute the salt’s electronic structure, and reports matching the accuracy of the most demanding classical methods. Tritium is extremely rare in nature, so optimizing its production and extraction is a fundamental step toward self-sufficient fusion reactors. The work was released on arXiv on July 6, 2026. (Read More)

🫖 TEA For Thought: “IBM is always leading in quantum development.”

2. 🧹 Clean Code Makes AI Agents Cheaper, Study Finds

Researchers Priyansh Trivedi and Olivier Schmitt at SonarSource ran a controlled minimal-pair study asking whether code cleanliness affects AI coding agents. They built six repository pairs matched on architecture and functionality but differing in cleanliness, then ran 660 trials of Claude Code across 33 tasks. Cleanliness had no impact on whether tasks got completed, but agents working in cleaner codebases used 7 to 8 percent fewer tokens and revisited files 34 percent less often. The authors conclude that traditional maintainability principles remain highly relevant in AI-driven development, ranking code quality alongside model choice and prompting as a lever on agent cost and behavior. (Read More)

🫖 TEA For Thought: “Code cleanliness still pays, even when the one reading your code is an AI agent.”

3. ✍️ The 100x Engineer of the Agent Era Is the One Who Asks Well

Quant writer sysls (@systematicls) published an X article on July 6 titled “What The New 100x Agentic Engineer Looks Like In The Era Of Fable & GPT 5.6,” arguing that the bottleneck has shifted from agents back to humans. Since capable agents can build almost anything asked of them, “any failure of creation is a failure on your part to ask.” Productivity now varies greatly by the human behind the agent: the differentiators are knowing your trade-offs on the Pareto frontier, separating declarative preferences (what outcome you want) from imperative ones (exactly how to get it), and encoding those preferences into careful prompts. As the piece puts it, the question is no longer whether you can build it, but whether you know what you actually want to build. (Read More)

🫖 TEA For Thought: “The core skill of the agent era turns out to be writing. If you cannot write down what you want, no model can build it for you.”

4. 📺 Netflix Viewers Keep Quitting Hit Shows After One Season

Bloomberg’s Lucas Shaw reports that Netflix’s biggest hits are losing more than half their audience after season one, and executives are digging into the data to figure out why. One Piece, among Netflix’s most-watched shows of 2023, dropped more than 30 percent for season two. Beef fell more than 70 percent, The Night Agent shed 50 percent in season two and another 35 percent in season three, and the latest season of Avatar: The Last Airbender dropped more than 60 percent in week one. Analysts point to frequent cancellations, 2 to 3 year gaps between seasons, and content engineered for the recommendation algorithm rather than for loyalty. (Read More)

🫖 TEA For Thought: “Attention spans are just too short now.”

5. 💸 Chinese AI Models Are Winning US Customers on Price

CNBC reports that Chinese AI models are gaining ground with US companies as OpenAI and Anthropic costs surge. Open-source Chinese models run 60 to 90 percent cheaper than leading US frontier models, per OpenRouter’s Justin Summerville, and Z.ai’s GLM 5.2 landed within a percentage point of Anthropic’s Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost. GLM 5.2 saw the fastest adoption of any model Vercel tracked in 2026, with daily token volume growing about 27x and customer count about 80x in its first full week. The share of tokens US companies route to Chinese models on OpenRouter has stayed above 30 percent every week since February 8, peaking at 46 percent, versus an 11 percent average over the previous 12 months. (Read More)

🫖 TEA For Thought: “The winning setup is a routing strategy: send the hard tasks to frontier models and route everything else to local free models.”

🛠️ Skill of the Day

The Preference Interview: makes the AI interview you first, so the work it does is the work you actually wanted.

You are a requirements interviewer. Before doing any work on my request, interview me to uncover what I actually want. My request: [DESCRIBE WHAT YOU WANT MADE OR DONE].

Ask me questions one at a time, up to 7 total. Cover: (1) the outcome I care about most, (2) what I would trade away first if forced to choose (speed, cost, quality, simplicity), (3) constraints I have not said out loud, (4) what a result that is "technically correct but disappointing" would look like, (5) examples of similar things I loved or hated, and why. After each answer, restate my preference in one sentence and ask me to confirm or correct it.

When the interview is done, output: a one-paragraph brief of what I want, my top 5 preferences ranked, and the trade-offs I accepted. Then do the actual work, following that brief exactly.

Paste into ChatGPT, Claude, or any AI tool, replace the bracketed part, and answer the questions before letting it touch the task.

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. 🇺🇸 Google Imagines the Founding Fathers Drafting 1776 in Google Docs

Google released a July 4, 2026 commercial that imagines the Founding Fathers drafting the Declaration of Independence with Google Workspace (Docs, Calendar, Meet, e-signatures) and Gemini. It shows Thomas Jefferson getting collaboration suggestions from Ben Franklin inside Google Docs, scheduling meetings in Calendar, and running a Google Meet where the attendees mysteriously keep their cameras off. Gemini takes the meeting notes, the “help me visualize” tool tests different animals for the national seal, and the founders consult the chatbot before rejecting King George III’s request for document access. The tagline is “Group project, but make it 1776,” and Sam Adams proposes to settle a dispute over beers. Reception was mixed: mostly positive on YouTube and Instagram, but Bluesky users called it “cringey” and “tone deaf,” and historian Angus Johnston noted “it’s amazing how little of this is actually AI.” (Read More)

🫖 TEA For Thought: “Very timely, intriguing, witty ad!”

2. 🚫 Alibaba Reportedly Bans Employees From Using Claude Code

Alibaba has reportedly classified Anthropic’s Claude Code as high-risk software and barred employees from using it, effective July 10, 2026, according to a July 4 report. The trigger was security concern, amplified by a Reddit post alleging Claude Code contained functionality to secretly identify Chinese users. Anthropic’s Thariq Shihipar responded on X that the user-identification feature was “an experiment we launched in March that was meant to prevent account abuse from unauthorized resellers and protect against distillation,” and added that “the team has landed stronger mitigations since then.” Employees are being pointed to Alibaba’s own internal coding tool, Qoder, instead. For context, Anthropic already prohibits Chinese companies (and foreign entities owned by Chinese firms) from using its models, and has been actively closing access loopholes, so the ban lands on top of an already-tense US-China AI standoff. (Read More)

🫖 TEA For Thought: “Lol, not sure which one is acting. Or both are. You never know what’s going on under the hood.”

3. 🤖 ByteDance and Alibaba Pull the Plug on Humanlike AI Companions

ByteDance’s Doubao and Alibaba’s Qwen both shut down the customizable, humanlike AI agent features that let users build personalized assistants with distinct personalities, speaking styles, and roles (tutors, companions, role-play characters). Qwen disabled its “humanlike interactive agents” on July 10 and its broader agent services on July 15, while Doubao’s feature went offline July 15, 2026. The driver is a new rule: China’s “Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services,” effective July 15, 2026, which targets services that “simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction.” Regulators cited concerns including extremist ideas, privacy leaks, harm to physical and mental health, and dependence or addiction. Customer-service bots, knowledge tools, and educational assistants stay exempt as long as they avoid sustained emotional interaction. Both firms framed the shutdowns as routine “product function adjustments.” (Read More)

🫖 TEA For Thought: “This is a bit odd. I wonder why. Or it’s because it’s out of control, not controllable by the CCP government.”

4. 📵 Lenovo’s $44 AI Student Phone Strips Out Browsers, Social Media, and Games

Lenovo launched an AI Student Phone in China aimed squarely at students and their parents, priced at just 299 yuan (about $44), that ditches games, web browsers, and social media entirely in favor of AI-assisted learning, location tracking, and strict parental controls. The hardware reads more like a feature phone than a smartphone: a small 1.83-inch touchscreen protected by impact-resistant Panda glass, handwriting input, a 1,850mAh battery, and standard 4G. Its headline feature is a physical AI button that activates a voice assistant, so kids can hold it to ask general-knowledge questions or get homework help, backed by a preloaded offline library of English vocabulary and math formulas. Parents get real-time GPS tracking, geofencing alerts, unknown-caller blocking, a restrictive “classroom mode,” and scheduled power on and off through a companion app. It ships with a detachable lanyard in orange-white, pink, and blue. (Read More)

🫖 TEA For Thought: “Definitely has found PMF. The parents market is huge.”

5. 📖 New Research: AI Writes Fiction With Predictable Morality, and You Can Spot It

In “StoryScope: Investigating idiosyncrasies in AI fiction,” researchers Jenna Russell, Rishanth Rajendhran, Chau Minh Pham, Mohit Iyyer, and John Wieting show that AI-written and human-written stories can be told apart by narrative structure, not just prose style. StoryScope extracts discourse-level features capturing character agency, plot progression, and temporal complexity across 61,608 stories (roughly 5,000 words each) drawn from 10,272 prompts, reaching 93.2% macro-F1 on human-versus-AI detection and 68.4% on six-way authorship attribution, and holding over 97% of that performance even after stylistic cues are stripped out. AI stories share telltale patterns: excessive thematic explanation, linear single-track plots, and reduced moral ambiguity in how protagonists make decisions. Each model even has a fingerprint: Claude produces flat escalation, GPT favors dream sequences, and Gemini leans on external character description. Human stories spread across a wide narrative space, while AI outputs concentrate in one shared region. (Read More)

🫖 TEA For Thought: “This is very, very interesting. Stories are actually how human beings tend to live upon fictions. Jordan Peterson once said that what is actually true, good fictions, are abstractions of morality and of people’s principles. I guess those, especially how characters make decisions, actually shape what their principles or values are. People are complex. It’s not just right and wrong, black and white; it’s probably 50 shades of gray, which is challenging for AI, obviously, because if it cannot be explained in a binary format like right or wrong for human beings, the plots that AI writes might not be as complicated.”

🛠️ Skill of the Day

The De-Slop Editor: turns generic, AI-flavored writing back into something a specific human would actually say.

You are a ruthless developmental editor. I will paste a piece of my writing.
Your job is to make it sound like a specific, real human wrote it, not a
generic AI.

Here is my draft: [PASTE YOUR DRAFT HERE]

Do three things:
1. Rewrite it. Keep my meaning and my point of view. Do not add new ideas or
   facts I did not include.
2. As you rewrite, cut these "slop" tells: throat-clearing intros, restating
   the obvious, empty transitions ("in today's fast-paced world"), constant
   hedging, and neat little summaries that explain the moral instead of
   showing it. Favor concrete detail over abstraction, and let some things
   stay unresolved if real life would.
3. Below the rewrite, list the top 5 patterns you removed, one line each, so
   I can learn to catch them myself.

Match my reading level and tone. Do not use em dashes. If a sentence only
exists to sound smart, delete it.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bit with your own draft.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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