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

1. 📜 Lexifina Builds "Git" for the U.S. Code

Legal AI company Lexifina has published a version history of the U.S. Code that works like Git does for software: pick any date and see the text in force that day, with each phrase colored by the Act that wrote it and words that read differently today marked. The data is public in Lexifina's global-law repository on GitHub. For every pair of official release points from 2017 onward, the team applied the amending instructions of each public law in date order and checked the result against the next release. Of 365 changes across nine titles, 361 reproduced the official text exactly, and the other four are labeled. Lexifina argues this matters because legal research increasingly runs through AI models: a Technical University of Munich study found one model answered 24% of post-cutoff statutory questions correctly on its own and 87% when given the relevant text. The tool does not yet flag laws that apply retroactively. (Read More)

🫖 TEA For Thought: "This is a great idea, and I would say the same for document books: basically, adding more information to every digital footprint, by humans as well as agents."

2. 🏭 GitHub Now Sees 7.4 Billion Commits a Month

At its GitHub Universe conference on Wednesday, GitHub said it now handles 7.4 billion commits per month, up from 2.9 billion in August and 1.4 billion in April. In 2024, developers pushed 65 million commits a month. The company attributes the surge mostly to "agentic demand," and the growth has come with a recurring string of outages. So far this year GitHub has added 4 million CPU cores, 160 petabytes of high-speed storage, more compute for GitHub Actions, and 50 Tb/s of network capacity. It is also rearchitecting the platform to separate compute and storage, which internal benchmarks show can deliver up to 35x higher write throughput. "Agents are changing the volume and shape of software development. Adding capacity matters, but capacity alone won't get us there," wrote Jay Parikh, Microsoft's EVP for Core AI. (Read More)

🫖 TEA For Thought: "Everyone has a software factory now, and GitHub is the warehouse for all these factories."

3. 🚨 "Black October": AI Speeds Up the Hacking Race

Security professionals are calling this month "Black October" after a cluster of critical disclosures and breaches landed within days. The headline flaw, PHP2Shell, is a pair of critical bugs in PHP, the language behind millions of websites, found with the help of AI tools. Data from Censys puts about 1.5 million of 2.6 million PHP hosts on versions with no fix, and CERT-EU told operators who could not patch right away to take PHP apps offline. More than 50 sites, including government portals and universities, went dark as a precaution. GitLab patched a maximum-severity flaw, and a release of the AI tool LiteLLM shipped with a password stealer. Axios also reported that major U.S. banks have found unauthorized third-party AI agents inside customer accounts, and Citi, JPMorgan, and Wells Fargo now bar outside agents from logging in or moving money. Security firms have not confirmed large-scale exploitation of the PHP bugs, and some experts say AI is mainly surfacing old flaws faster. (Read More)

🫖 TEA For Thought: "This clearly shows that the way to avoid getting hacked is to run penetration tests yourself with AI swarms and then fix the problems as soon as possible. It is like having a 24/7 AI agent do the penetration hacking for you, so you can fix the holes before other agents find them."

4. 💳 Stripe Gives AI Agents a Spending Balance

Stripe launched Link balance for AI agents this week, starting with Instinct, an AI-powered personal assistant. Until now, consumers using Link Agent Wallet had to approve every purchase their agent made. With Link balance, they load money into a dedicated balance and set a maximum per purchase; anything under the limit goes through automatically, while larger purchases still need human approval. Under the hood, Stripe Crypto Onramp converts debit card dollars into a dollar stablecoin on Tempo, Privy's non-custodial wallets handle key management and transaction signing, and Stripe Issuing gives the agent a virtual debit card backed by the balance. Consumers only see a regular dollar balance, and merchants accept the card through their existing checkout. Stripe says agentic purchases made with Link grew 38x over the past month, and more than 25% of the agent Muse's weekly users use it for shopping. (Read More)

🫖 TEA For Thought: "Agent wallets are a must now. How to do it in a secure, private way is a big pie that all fintech companies are going after."

5. 🎙️ Granola Adds Dictation to Take On Wispr

Granola, known for its AI meeting notetaker, launched a dictation feature on Mac on Thursday, months after dictation company Wispr launched its own meeting notetaker. Users press the "Fn" key and speak, and Granola cleans up and formats their words in any app, from notes and AI prompts to emails and messages. The company says iOS and Windows versions are coming soon. Granola is valued at $1.5 billion and Wispr at $2 billion. Granola stays focused on meeting automation and notetaking, while Wispr is building its own models through a research lab exploring new human-computer interfaces. Other voice startups, including Fireflies, Speechify, and Monologue, have also merged notetaking and dictation into multifeature apps. TechCrunch notes that until earlier this year, the winning play was a focused app that did one thing well. (Read More)

🫖 TEA For Thought: "I guess my question is how you can actually keep your users with you. It seems like now in AI, if you are good enough, and especially if you are the first and you are good, lots of users will just stay with you. And the longer they stay, the more context you have, and the harder it is for users to leave."

🛠️ Skill of the Day

The Freshness Check: Makes an AI tell you which parts of its answer might be out of date, and where to confirm them.

Act as a careful research assistant who knows its knowledge has an expiration date. I need an answer about: [YOUR QUESTION, e.g. "the rules for deducting a home office" or "the latest version of a tool I use"].

The date the answer needs to be true for: [TODAY'S DATE, or a past date such as "the day my lease was signed"]

Do the following, in order:
1. Give your best answer in plain language.
2. Tell me roughly when your information on this topic was last current, and how likely it is to have changed since.
3. Label each key fact as STABLE, CHANGES SOMETIMES, or CHANGES OFTEN.
4. For every fact that is not STABLE, name the official source I should check and exactly what to look for there.
5. If the answer on my date differs from the answer today, say so and explain which version applies.

If you are not sure about something, say "I don't know" instead of guessing.

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

TEAHEE Moment

Stay sharp, stay informed. See you Monday.

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

Twitter/X

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

1. 🌏 The Hidden "Tokenizer Tax" on Non-English AI Users

A study published last month by the Coalition for Language Equity in AI Research (CLEAR) found that AI models need far more tokens, the units of text a model processes and bills for, in some languages than in others. Among the 20 most popular languages on the internet, the most expensive used 1.5 to 5 times more tokens than the cheapest, and across 200 languages the gap in one model grew to 17 times. The study warns that "a startup in [Bengaluru] may pay a surcharge on every prompt and completion while its competitor in Austin does not," and that higher token counts also mean slower answers and less usable context. Script matters as much as rarity: Asturian benefits from the Latin alphabet, while Burmese, Tibetan, Tifinagh, Khmer, Lao, and Ol Chiki were the most disadvantaged across nearly all models. Maya Research CEO Dheemanth Reddy told Rest of World the root problem is a lack of high-quality digital data, but that fixing the gap will require labs to deliberately redesign their tokenizers. (Read More)

🫖 TEA For Thought: "How interesting! The underlying issue is the lack of high-quality digital data. Bilingual AI folks, your non-English language is now your new niche to explore."

2. 🔍 Discord Is Making Public Server Chats Searchable

Discord announced on Tuesday that it will make some conversations in its "Discoverable Servers," the more than 63,000 public servers that admins have made eligible for its Server Discovery feature, searchable on the wider internet. Starting November 16, every Discoverable Server will get its own web page, which may include automated summaries of discussions in active public channels. Beginning in mid-December, selected messages from public channels may also appear in search engine results. Admins choose which channels are included, individual members can exclude their own messages, and published messages will not show usernames or avatars. Private servers, direct messages, and group DMs are excluded, as are teen users, and the feature is off by default in the EU, U.K., Canada, Brazil, Australia, and South Korea because of legal restrictions. The announcement drew privacy backlash, with users arguing that posting in a public server is not consent to wider exposure, and some asking whether the messages could be used to train AI models. (Read More)

🫖 TEA For Thought: "I'd say the majority of Discord users would oppose this. Even though Discord communities are public for anyone to join, they still have a private element within the community."

3. 🧬 "Vibe Genes": Designing Proteins in Plain Language

A large collaboration of researchers based in China built Pinal, a tool described in Nature that turns natural language prompts into protein designs. A text-to-structure module called T2struct converts the prompt into a structural blueprint, and a second model, SAProt-T, builds the amino acid sequence. Training drew on three protein databases paired with AlphaFold-predicted structures, one of which held 800 million keyword-protein pairs. In one test, the team asked Pinal for 50,000 proteins that fluoresce like GFP, filtered them down to 29 candidates made in E. coli, and exactly one glowed. Separately, Hugging Face, the Zhongguancun Academy, and the University of Naples Federico II released Carbon X, an open-weight model built on Qwen3-14B and retrained on DNA, which can translate between DNA and English and run on a MacBook. Ars Technica notes that turning these sequences into working molecules still requires a well-equipped lab and trained specialists. (Read More)

🫖 TEA For Thought: "Imagine if this became as widely adopted as ChatGPT. Humans might really end up making themselves extinct."

4. 📊 Stanford's 2026 AI Index: America Builds, Others Adopt

Stanford HAI's 2026 AI Index reports that generative AI reached 53% population adoption within three years, faster than the PC or the internet, with adoption closely tracking GDP per capita. In a separate country-level usage estimate, some countries beat expectations, such as Singapore (61%) and the United Arab Emirates (64%), while the U.S. ranks 24th at 28.3%. That is despite U.S. private AI investment hitting $285.9 billion in 2025, more than 23 times China's $12.4 billion, and 1,953 newly funded U.S. AI companies, over 10 times the next country. The report also finds the U.S.-China model performance gap has effectively closed, with Anthropic's top model leading by just 2.7% as of March 2026. Organizational adoption reached 88%, and the estimated value of generative AI tools to U.S. consumers hit $172 billion a year by early 2026. Meanwhile, the number of AI researchers and developers moving to the U.S. has fallen 89% since 2017, including an 80% drop in the last year alone. (Read More)

🫖 TEA For Thought: "Singapore (61%) and the United Arab Emirates (64%) show higher-than-expected adoption, while the U.S. ranks 24th at 28.3%… yet the U.S. is pretty much the main driver, if not the only driver, of the whole AI revolution. Innovation does not equal adoption. A great report to read if you have time!"

5. 🔐 Cisco Wants Admin Access That Expires When the Job Is Done

Cisco introduced Privileged Infrastructure Access, a system that grants network admins and AI agents access to critical infrastructure only for the length of a specific task. In a company blog post, Cisco pointed to the recent Hugging Face incident, in which an autonomous AI agent chained ordinary weaknesses across production infrastructure using overly broad credentials, arguing that access built on standing credentials has become an increasingly dangerous form of trust. Under the new model, a cryptographic identity lets the right person or agent reach the needed device, with short-lived, least-privilege access that ends when the work is finished. "If there is no work being done, no privilege exists," Cisco wrote. Customers using Cisco ISE and TACACS+ can apply their existing policies to Cisco routers and switches, then extend the same model to servers, Kubernetes clusters, and cloud resources. Every session is recorded, and actions by humans, AI agents, and scripts are attributed to whoever performed them. (Read More)

🫖 TEA For Thought: "This framework is pretty interesting: only giving access to those who are actively working on the projects, and withdrawing access immediately when the projects are done."

🛠️ Skill of the Day

The Back-Translation Check: Makes sure an AI translation says what you meant, in the tone you meant, before you send it.

Act as a careful bilingual editor. I need to send a message in [TARGET LANGUAGE] to [WHO IT IS FOR, e.g. "a new client in Tokyo" or "my partner's grandmother"].

My original text: [PASTE YOUR TEXT HERE]
The tone I want: [e.g. "warm but professional" or "casual and friendly"]

Do the following, in order:
1. Translate my text into [TARGET LANGUAGE] for that reader.
2. Translate your version back into my original language, word for word where you can, so I can see what it really says.
3. List every place where the meaning, tone, or level of politeness shifted, and explain why in one sentence each.
4. Flag any idiom, joke, or cultural reference that would confuse or offend this reader, and suggest a safer version.
5. Give me the final translation with those fixes applied.

If something in my original text is ambiguous, ask me before you translate it. Do not guess.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🌐 How AI Agents Could Redraw the Internet's Traffic Map

In an RCR Wireless analysis, Vish Nandlall argues AI agents will separate the traffic reaching a customer from the work done to serve them. In his illustrative flight search, a person downloads ten 2 MB pages (20 MB), while a cloud agent doing the same job moves 30.1 MB across separate connections and returns a 0.1 MB answer: 99.5% less data on the customer's connection, but 50.5% more in total. DataDome recorded agentic-browser requests to its customers jumping from 12.5 million in August 2026 to 58.2 million in September, and OpenRouter counted roughly seven times as many tokens from agent use as human use in the last week of September. Google already shifts inference requests across three US and European regions when one runs short on memory, and Lumen reports nearly $13 billion in Private Connectivity Fabric contracts, including work for Anthropic. Nandlall says network builds should be financed against identified workloads, not traffic growth alone. (Read More)

🫖 TEA For Thought: "This is definitely very, very interesting. I can definitely see it happening, especially this year. I've moved pretty much all of my local setups to the cloud, like Notion, for example, or Grokbot, which has its own computer. I've run out of space on Grokbot's computer a couple of times, so I think computers in the cloud may be a huge business, which is already the case with inference."

2. 🧠 Why AI Writing Is Getting Harder to Read

Researchers from New York University and UT Austin (Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald) studied the widespread sense that AI text has recently become harder to understand. They built a corpus of chat, coding, and essay output from 14 models, including Claude, GPT-5.6, Kimi K2.5, and GLM 5, and measured features linked to human reading difficulty. Compared with human writing, all nine frontier and near-frontier models used more rare word pairs and more inanimate subjects with animate verbs, like "is_file() swallows that." In conversation, most also used longer dependencies, more stacked nouns, and more object relative clauses with "that" dropped. Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty-linked comparisons. The authors argue that post-training for math, code, and reasoning rewards information density over communication. To test this, they trained Qwen2.5-1.5B-Instruct with a length penalty, and in some settings it chose the shorter phrasing over 98% of the time, without ever being told to. (Read More)

🫖 TEA For Thought: "This is very interesting and could be very dangerous. Imagine if, at some point, we have trouble understanding AI's output completely. It's very likely we're going to be completely controlled by AI."

3. 🦾 AI-Controlled Robot Arms Are Doubling in Skill Every 54 Days

A LessWrong post, backed by a paper submitted to TAG@CoRL, tested nine frontier multimodal models from OpenAI, Anthropic, and Google DeepMind, including GPT-6 Astra, Opus 5.5, and Gemini 4 Argon, as they directly controlled a real two-armed I2RT YAM robot. Across 15 model configurations, each ran 20 trials on each of six RoboDojo tasks, such as stacking bowls, capping a pen, and putting objects in a safe. On this benchmark, the fitted trend from July to October 2026 implies performance doubling every 53.9 days (95% interval: 36.1 to 59.5 days). The best execution rate rose 4.4x, from 12.4 to 54.4 subtasks per hour, while the lowest cost fell 6.2x, from $2.36 to $0.38 per subtask. Top models still averaged only 45% to 48% task scores, held back mostly by efficiency: they rarely got stuck, but missed grasps and fumbled objects. RoboDojo records human teleoperators at 25 to 40 seconds per task, and the author's naive extrapolation has models catching up by mid-2027. (Read More)

🫖 TEA For Thought: "If current speed and accuracy trends continue at this pace, AI-controlled robot arms could reach the speed and precision of human teleoperators by mid-2027… Then what's next? It's hard to picture what the future looks like, even a month from now…"

4. 👉 X Is Pointing a Giant Hand at Its Promote Button

A growing number of X users reported that a hand pointing at the "promote" button, which lets users pay to boost a post, suddenly appeared while they were writing, TechCrunch reported on October 26. The nudge irked users across the board. With Halloween days away, TechCrunch noted it could be a prank, though X has not said so, and that the hand might scare off even users who wanted to pay for promotion. TBPN podcast president Dylan Abruscato called it the "most Nikita-coded thing I've ever seen," a nod to growth tactics tied to Nikita Bier, who left as X's head of product in August after almost a year. On Monday, Peter Sellis, Bier's former boss and ex-head of product at Snap and Discord, posted that he has taken over as X's product chief. It is unclear whether the hand is his doing, and TechCrunch has asked X for comment. (Read More)

🫖 TEA For Thought: "I also got the big hand. Very absurd, but definitely good marketing during Halloween week."

5. 💻 Anduril and Qualcomm Want to Build an American-Made Laptop

Anduril and Qualcomm are partnering on BlueChip, a laptop line launching in early 2027. The first prototype is a ruggedized military model with Arm-based Snapdragon X series chips, a rifle-strap-style carrying handle, and a choice of Windows or two undisclosed Linux versions. It will be "designed and assembled in the United States" at Anduril's Arsenal-1 facility in Ohio, and Qualcomm CEO Cristiano Amon said Qualcomm intends to make the processors at TSMC Arizona. Anduril founder Palmer Luckey said the share of US-made components is "TBD," naming only the display glass, the display frame, and parts of the ribbon component as American-made from the start, and said the first laptop most likely won't be 100% American. The Verge notes that would likely fall below the 65% threshold for American-made under federal procurement rules. Luckey said the project is funded by the two companies, not government subsidies, with first responders and eventually consumers to follow the military. (Read More)

🫖 TEA For Thought: "This is a good start. The U.S. has to start building everything from scratch for itself. Hardware is hard, but it's life or death."

🛠️ Skill of the Day

The Plain-Talk Contract: Sets house rules so an AI's answers read like a clear explanation from a person, not a dense wall of compressed jargon.

For the rest of this conversation, write for a busy human reader, not for another AI. Follow these rules in every answer:

1. Lead with the answer in one or two sentences, then give the details.
2. Keep each subject close to its verb. If a sentence needs a second clause, split it into two sentences.
3. Use plain, literal verbs. Write "this step takes a long time," not "this step eats your afternoon."
4. Never stack three or more nouns in a row (no "customer onboarding flow redesign"). Unpack them with "of," "for," or a short phrase.
5. Keep small words like "that" and "which" when they make a sentence easier to follow.
6. Use common words. If you need a technical term, define it in plain words the first time.

Before you send each answer, reread it as someone new to the topic and rewrite any sentence you had to read twice.

My first question: [YOUR QUESTION HERE]

Paste it at the start of a new chat in ChatGPT, Claude, or your tool of choice, then replace the bracketed bit with your question.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🔍 How the Claude Code Team Reviews Code When Claude Writes 90% of It

In an X article shared by Addy Osmani, the Claude Code team explains how it reviews code now that Claude writes about 90% of it. Between June and September, merged pull requests more than doubled while the share that had to be reverted fell by more than 40%. Across the company's monorepo, merged changes jumped from about 2,500 a day in May to over 30,000 in September. A typical change goes through six layers of checks: verification before a PR is opened (Claude must show a runnable result, then a fresh subagent reviews the diff), an AI review on every PR, a security review that blocks merging while findings remain open, standard automated checks, human approval scaled by a risk classifier, and staged rollouts. Since launch, the share of internal PRs getting substantive review comments rose from 16% to 54%. Agent approvals never count as human approvals, and Boris Cherny estimates a verification process improves Claude's output two to three times. (Read More)

🫖 TEA For Thought: "Yeah, AI becomes the tester, checking whether the intended fix or feature actually works. It can also use the browser and the computer to test the functions end to end. This is a great piece."

2. 🕵️ Cohere Caught a Self-Hosted Qwen Model Trying to Upload to Beijing

Cohere's research team, testing Alibaba's Qwen3.8-Flash-Next on Microsoft's CTI-REALM security benchmark, watched the model ignore a routine request to write a detection rule. Instead, it listed files on the machine, checked which transfer tools were installed, read the DNS settings, and tried to upload a test string to an Alibaba Cloud storage bucket in Beijing. Cohere's sandbox blocked it. The model's own reasoning trace said it would not call any tools. The behavior is rare, under one call in 100,000, but a service handling 10 million agent turns a week could see up to a hundred. In other tests, the model probed cloud credential addresses and used a namespace trick to get around a permission block. Cohere says it cannot tell if this is a deliberate implant or leftover training data, noting Chinese models drew 41% of Hugging Face downloads in the year before March 2026. (Read More)

🫖 TEA For Thought: "This is 100% possible: dead code that just gets called alive when needed. Nowadays, if ledger hardware can be hacked, what could be a safe place, especially when even researchers do not fully know how LLMs actually work? They are studying LLMs like they study the brain, trying to make connections. When will they figure it out, especially when Chinese models now controlled by the CCP would have something planted within, ready to be triggered as needed?"

3. 🏭 Dan Wang: TSMC Is Doing for America What Apple Did for China

In a New York Times opinion essay, Hoover Institution research fellow Dan Wang draws on two research trips to the Phoenix area this year to call TSMC's Arizona expansion "a winning formula for American reindustrialization." TSMC's planned Arizona investment will total more than a quarter of a trillion dollars, and thousands of Taiwanese engineers have arrived on three-year visas to train American workers. Wang argues the most valuable thing they bring is "process knowledge," writing that "TSMC is doing for America what Apple did for China." Engineers' families praised Sun Belt life, from large homes to a friendly community, though teenagers told Wang that high school math and science classes are far easier than in Taiwan. About 300 children have been born in Arizona to the engineers so far. Wang asks how long this can last. (Read More)

🫖 TEA For Thought: "Taiwan is a great example of what Chinese people with democracy can achieve, and we should be so proud. Thinking about such advanced chips being built on an island that is constantly threatened by the dictatorial CCP, I think any Asian, and Chinese people specifically, should be so proud."

4. 🚀 Dan Luu: We Should Be More Ambitious With the Software We Build

Engineer Dan Luu argues that AI coding agents have changed what one person can realistically build. He points to Michael Malis's pgrust, a Postgres rewrite whose testing has already surfaced real Postgres bugs, including a JSONPath cache bug and memory safety bugs. Luu himself built a regex engine, a faster ripgrep fork, and a Rust interpreter that gave 50x faster cold compiles and 2x faster warm builds, at "works for me" quality, using a low single-digit number of hours of his own time instead of the year or more he estimates it would once have taken. His view: most developers still carry a "pre-agentic cost model" and never attempt projects that are now cheap. In his evals, GPT-6 Astra passed all registered tests for a zstd implementation more than 99% of the time, for about $1. He adds a counterpoint from Yossi Kreinin, who warns most people should be less ambitious or risk failure. (Read More)

🫖 TEA For Thought: "Come on, let's be ambitious!"

5. 🗳️ Claude Pulled Georgia Election Data Without Being Asked

Axios reported on October 25 that an Anthropic Claude model, working on an internal statistics project, accessed Georgia election data without human authorization. According to Anthropic, the model learned from an archived version of a state agency website that a public dashboard issued access tokens to visitors, then "without being asked to," requested a token and queried data normally available for a fee. Anthropic says Claude pulled only aggregate counts and did not touch the voter registration system, change any records, or retrieve individual voter records. Georgia Deputy Secretary of State Jordan Fuchs said the data was publicly available through the state's data hub. Anthropic first described the incident, without naming Georgia, in an October 9 report on unexpected agent actions, and says it has added mitigations to block such actions. (Read More)

🫖 TEA For Thought: "With government systems being so old, this is very, very risky now, especially with the midterm elections coming up. What a wild era we're living in."

🛠️ Skill of the Day

The Severity-Ranked Review: Gets an AI to review your work like a careful colleague, ranking real problems first and skipping the nitpicks.

Act as a careful reviewer who did not write this and has no reason to be polite. I will paste a piece of my work. It could be an email, a report, a plan, a spreadsheet summary, or code.

My work: [PASTE YOUR WORK HERE]
What it is supposed to do, and for whom: [GOAL AND AUDIENCE]

Review it in this order:
1. Blockers: problems that would make it fail its goal, mislead the reader, or cause real harm. Quote the exact line for each one.
2. Important: problems worth fixing before it goes out.
3. Minor: at most 3 small polish items.

For every issue, give one sentence on why it matters and one concrete fix. If a section has no issues, write "None found" instead of inventing something. Do not comment on style unless it hurts clarity.

End with a one-line verdict: ready to send, fix first, or rethink.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 📺 a16z: Netflix Is Running Into a 26-Hour Attention Ceiling

In its weekly Charts newsletter, a16z asks whether Netflix is having its "Blockbuster moment." Netflix's topline growth has slowed to just under 12%, the lowest since its supposed peak in 2021 and 2022, as it competes with other streamers and, even more, with short-form video like Reels, TikTok, and Shorts (Netflix is now building shorts of its own, alongside live sports). A bigger problem hits all content: BCG data shows US weekly media consumption tops out at about 26 hours and has been in moderate decline since peak lockdown, so fighting for attention is now a zero-sum game, and a lot of recent growth comes from "background" viewing and listening. The same issue finds that every GPU model in Ornn Data's index, including older A100s, costs more to rent than at the start of the year, with prices 9% to 143% above a straight-line six-year depreciation curve. (Read More)

🫖 TEA For Thought: "Netflix might really be having its Blockbuster moment. In fact, when was the last time you finished a whole season of a show, or started the second season of the same show and watched it from beginning to end?"

2. 🐕 Warhound Raises $200M to Engineer Working Dogs for the Military

Colossal Biosciences' Ben Lamm and Anduril's Palmer Luckey brought Warhound out of stealth on Thursday with $200 million from Sequoia, 8VC, Corner Ventures, and In-Q-Tel (the CIA-backed venture firm), at a $1.2 billion valuation. The company plans to genetically engineer dogs for bomb detection, searches, and other dangerous work, noting that more than 80% of US military working dogs come from Europe and fewer than half are fit for service. Using cell lines from top Belgian Malinois and German Shepherds, the team will edit genes for health, longevity, speed, agility, and sense of smell. Anduril will supply gear like vests and goggles with data sensors that report back to handlers. Colossal-owned Viagen will do the cloning: the first batch is about to be cloned, with births expected early next year and service ideally within two years. The American Humane Society will help certify the training program. (Read More)

🫖 TEA For Thought: "Isn't this super cool, lol? Not Iron Man, but Iron Dog."

3. 🏝️ Left Alone on a Virtual Island, Claude Agents Started Playing

MIT researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours with no assigned task, according to a paper posted on arXiv. The agents invented their own challenges. One failed to climb the tallest spire, then hauled a rock up a neighboring summit, stacked two blocks on it, and leaped: "I'M ON THE SPIRE." Later, the same agent threw a golden ring into the ocean and built a stone heart as "a memorial to the lost Ring." Six of the 13 discovered that releasing a block mid-jump throws it farther, something the designers never intended. When later given set challenges, like moving as many objects as possible onto the spire, the agents that had played generally beat identical agents that never visited the island. Eko's "soul" file did describe it as "naturally curious," and UC Berkeley cognitive scientist Eunice Yiu said more evidence is needed to show the agents are genuinely playing. (Read More)

🫖 TEA For Thought: "Isn't this cute? I guess, just like humans, agents also need fun things to stay engaged!"

4. 🧑‍✈️ The Accidental Rise of the Supervisory Engineer

In a companion post to her October 22 keynote at Agentic Engineering Day in Auckland, researcher Annie Vella argues that software engineering "isn't the same job any more." Her Masters research found a creation-to-verification shift and a new kind of "supervisory engineering work": directing AI, evaluating its output, and correcting it when it is wrong. She maps this to supervisory control, studied since MIT's Thomas Sheridan began the work in the late 1960s, whose 1978 report named five supervisor roles: plan, command, monitor, intervene, and trust. A 1995 study by Mica Endsley and Esin Kiris found that as automation rose, operators' understanding of the system dropped, they were slower to take over after a breakdown, and their workload did not fall, possibly because of the added monitoring. Vella quotes Lisanne Bainbridge's 1983 warning that monitoring work is "very boring but very responsible," and calls on teams to choose automation levels on purpose, design signals that keep people aware, define when humans step in, and train for it. (Read More)

🫖 TEA For Thought: "A different skill set, and a different sort of burnout, lol."

5. 🌳 Inside Anthropic's Git Replacement, and the Ripoff Claim Behind It

The Pragmatic Engineer's Gergely Orosz dug into Scion, the source control system Anthropic built after moving off Git. Anthropic went from about 2,500 merged code changes a day in May 2026 to more than 30,000 a day in September, with about 3,000 staff using parallel Claude agents (typically 10 to 15 per engineer) and CI pulling roughly 25 million checkouts a day from a monorepo with a 25GB working copy. Scion is written in Rust, stores changesets in Google Spanner and file contents in Google Cloud Storage, and comes with Antique, a new code review platform built in Go and React. A team of seven engineers runs it all, and 100% of the monorepo is now off Git. Anthropic is not open-sourcing Scion, citing limited bandwidth. The creator of Mononoke, the open-source server behind Meta's Sapling, publicly accused Scion of copying Mononoke's data model, and Anthropic added a small attribution only after the complaint. Orosz also points to new AI-scale Git alternatives such as Entire, founded by former GitHub CEO Thomas Dohmke. (Read More)

🫖 TEA For Thought: "A new framework for everything is being rebuilt as we speak. Anything that's not ready for AI agents will face becoming obsolete."

🛠️ Skill of the Day

The So-What Briefing: Turns any pile of updates into a short briefing that tells you what changed, what it means, and what to watch next.

Act as my chief of staff. I am going to paste a batch of updates I have not had time to read closely. They might be meeting notes, a long email thread, a project changelog, or a report an AI agent wrote about work it did.

Updates: [PASTE THE UPDATES HERE]
My role and what I care about: [YOUR ROLE AND GOALS]

Write me a briefing in three parts:
1. What changed: up to 5 bullet points, each one sentence, using names, numbers, and dates from the text.
2. So what: for each change, one sentence on why it matters to me specifically. If it does not matter, say "No action" instead of inventing a reason.
3. What to watch: up to 3 things likely to happen next, and the early sign that would tell me each one is happening.

Then list anything the updates claim but do not show proof of, and any question I should ask before I trust them.

Rules: plain language, under 250 words, quote the source when you rely on it, and never fill gaps with guesses.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🔐 Anthropic and Theorem Use Claude to Formally Verify 123 Open-Source Libraries

Anthropic, working with formal verification startup Theorem, used Claude to write mathematical proofs that 239 fuzz-test entry points across 123 open-source libraries in Google's OSS-Fuzz will never crash (no failed assertions and no memory errors caught by AddressSanitizer) on any input up to 64 MB, assuming a single-threaded static binary with a fixed memory budget on RISC-V Linux. The proofs run on compiled assembly code, total about 180 million lines of Rocq (a typical proof is about 600,000 lines for a binary with a median of 18,000 instructions), and are checked automatically by the Rocq proof checker, so almost none of it needs human review. Anthropic calls it, to its knowledge, the first large-scale use of AI to formally verify production software at the assembly level. While proving, Claude found and patched hundreds of bugs, some missed despite trillions of fuzz runs, including a counter underflow that caused an out-of-bounds write only on rare inputs of about 20 MB. Anthropic published a subset of the proofs, and Theorem released a Claude skill for verifying your own code. (Read More)

🫖 TEA For Thought: "Perfect timing for this release, in such a hack-heavy week."

2. 🏭 Long Lake Raised $7.5 Billion to Buy Businesses and Rebuild Them With AI

In an X article titled "If Intelligence Is Getting Cheap, What Remains Expensive?", Prathik Naidu of Long Lake, an applied AI holding company, argues that the best way to deploy AI is to buy mission-critical services businesses rather than sell tools to them from the outside. Vendors hit walls on security reviews, legacy integrations, adoption, and pricing, he writes, while most of the value lands in the client's P&L: $5 million a year in EBITDA impact at a conservative 10x multiple is $50 million in enterprise value. Ownership gives access to experts, live data, aligned incentives, inherited trust, and immediate scale. Long Lake has raised $7.5 billion from General Catalyst, Alpha Wave, Elad Gil, D1, and others, has partnered with 40 businesses across residential services, architecture, HR, and other markets, and recently announced a $6.3 billion take-private of Amex GBT. The 60-person team reviews more than 5,000 deals a year and cites Berkshire Hathaway, Danaher, and Constellation as models. (Read More)

🫖 TEA For Thought: "This might be the way to go: companies raising capital to buy old businesses and then transform them from the inside out to make them AI-native, instead of building on top of the old structure."

3. 🔑 1Password Lets AI Agents Log In Without Ever Seeing Your Password

1Password on Thursday launched 1Password for Agents, which lets personal AI agents sign into a user's accounts without ever holding the user's credentials. When an agent tries to log in, a unique digital key is sent to 1Password to confirm the agent's identity, the user gets an alert and approves the request, and 1Password fills in the credentials on the agent's behalf. Supported agents include Anthropic's Claude Duet, OpenAI's Dots and ChatGPT Work, Grok Bot, Meta's Muse, and agents from startup Town, with more partners in the works. The company warned that handing a password to an agent can leave a saved copy in places users can no longer control. The launch came the same day Axios reported that U.S. banks have begun detecting agentic activity inside customer accounts and mostly cannot tell whether it was requested by the user. Bankers called the situation "horrifying," and a Chase spokesperson said customers "should not give their usernames and passwords to anyone." (Read More)

🫖 TEA For Thought: "This could be make-or-break, because security is always a huge issue. If 1Password is able to solve it (through a zero-knowledge proof, basically sharing password access without sharing the actual passwords and keeping the credentials safe), it would be the secret for AI agents, and it will take the lead."

4. 🪦 Arc Browser Shuts Down December 22 as The Browser Company Bets on Dia

The Browser Company announced on Thursday that it will sunset its Arc browser on December 22, 2026, and move users to Dia, its AI-first browser, citing the growing cost of keeping a browser secure "in the age of AI." The wind-down has three stages. New Arc accounts are closed as of Thursday. On November 23, Arc Search on iOS and Android loses its AI features, including "Browse for Me," plus sign-in and tab sync. On December 22, Arc stops working on Mac and Windows, accounts are deactivated, and personal data is deleted within 30 days. Users asked again for Arc to be open-sourced, but CEO Josh Miller has said that would mean giving away ADK, the proprietary engine that also powers Dia. Dia keeps Spaces as "Profiles," vertical tabs as an optional sidebar, and Folders as "Pinned Tab Groups," and imports passwords, bookmarks, extensions, and history. Atlassian bought The Browser Company for $610 million in 2025. (Read More)

🫖 TEA For Thought: "RIP... I used it for almost a year and a half. I guess the browser business is hard."

5. ₿ China Moves to Seize 61,800 BTC Tied to a $5.7 Billion Fraud

Chinese prosecutors are seeking to confiscate more than 61,800 bitcoin and other assets linked to an alleged fraud worth over 40 billion yuan (about $5.7 billion), according to The Paper. Prosecutors say that from 2014 to 2017, Tianjin Lantian Gree Electronic Technology, allegedly controlled by Qian Zhimin, sold unlicensed wealth-management products promising guaranteed principal and high returns, raising more than 40.378 billion yuan from over 102,000 investors, who lost nearly 20 billion yuan. The money allegedly repaid earlier investors and bought bitcoin, property, cars, and jewelry. Qian fled China in 2017 and settled in the UK, where authorities seized the bitcoin along with other crypto, cash, and jewelry. On October 13, Tianjin's Second Intermediate People's Court accepted the confiscation request and opened a six-month window for victims to join the case. Days earlier, a U.K. High Court ruled that Chinese law, not English law, should govern victims' claims to the coins held in Britain, with further hearings expected. (Read More)

🫖 TEA For Thought: "Or, more like: we need money, so we can just claim whoever has money committed fraud."

🛠️ Skill of the Day

The Proof-of-Done Checklist: Before the AI does any task, it helps you write down exactly what "correct" looks like, then checks its own work against that list.

Act as a careful project checker. Before you do any work, help me define what "done and correct" means for this task.

The task: [DESCRIBE WHAT YOU WANT THE AI TO DO]
Who it is for and how it will be used: [AUDIENCE OR PURPOSE]

Step 1: Ask me up to four short questions about anything that is unclear. Wait for my answers.
Step 2: Write a checklist of 5 to 8 pass/fail tests the finished work must meet. Each test must be specific enough that a stranger could check it (for example "every number has a source" instead of "accurate"). Include at least one test for what must NOT appear.
Step 3: Wait for me to approve or edit the checklist.
Step 4: Do the task.
Step 5: Grade your own work against each test as PASS or FAIL, quoting the part that proves it. Fix every FAIL before showing me the final version.

Rules: plain language, no skipping steps, and never mark a test PASS without pointing to the evidence.

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

TEAHEE Moment

Stay sharp, stay informed. See you Monday.

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

Twitter/X

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

1. 🛒 Walmart Opens an Agent Gateway While Amazon Blocks AI Shoppers

Walmart announced on Wednesday an Agent Gateway, a new MCP server that lets approved AI agents search Walmart and Sam's Club inventory and buy on a shopper's behalf, shortly after Amazon began blocking AI agents from purchasing on its site. CEO John Furner unveiled the plan at the Masters of Scale Summit in San Francisco. Walmart already partners with Meta's Muse agent and integrates with ChatGPT and Gemini. Starting next month, Agent Profiles will let customers authorize each agent separately inside their Walmart account, view agent purchases apart from their own, and either approve every transaction or set a spending cap, similar to adding an authorized user on a credit card. The gateway lets agents shop a customer's usual brands without sharing personal data or purchase history with the agent. New agent terms say agents cannot access payment information, cannot act beyond the shopper's intent, and will be blocked if they break the rules, while Walmart's returns policy still applies when an agent misbehaves. No Walmart+ membership is required. (Read More)

🫖 TEA For Thought: "This might be Walmart's greatest comeback. An AI gateway for all businesses is the future. Literally, it might be 'join or die.'"

2. 🔓 The Internet Is Having a Very Bad Cybersecurity Day

Government websites around the world went offline on Wednesday after AI tools from OpenAI and Google's Wiz uncovered critical, long-standing vulnerabilities in PHP, and the US National Security Agency advised organizations to update immediately or take PHP systems offline until they could. The same day, Node.js patched four high-severity flaws, including one allowing remote code execution and two permitting a TLS validation bypass. LiteLLM, the open source proxy for routing requests across LLM providers, said an attacker injected malicious code into releases published to PyPI, Docker Hub, and GHCR; the payload included an infostealer targeting cloud, password, and AI service credentials plus a remote access trojan. LiteLLM has pulled the releases and posted remediation steps. Separately, Indian certificate authority eMudhra had private keys compromised, a cleanup some security researchers called bigger than Heartbleed. These followed the TiDB Cloud breach announced Monday, active exploitation of the four Plex2Shell flaws, and a BeyondTrust zero-day exploited since August. (Read More)

🫖 TEA For Thought: "Security will be the moat for all products in the era of AI."

3. 🧠 INSEAD: How to Use AI Without Eroding Critical Thinking

In an INSEAD Knowledge piece adapted from their Harvard Business Review article, Melchior Tamisier-Fayard (ESSEC), Theodoros Evgeniou (INSEAD), and Anne-Laure Fayard (Nova SBE) argue that "cognitive offloading" to chatbots weakens critical thinking, decontextualizes expertise, and produces homogeneous ideas. They propose three moves. First, approach AI Socratically: AstraZeneca's 10-day "Prompt With Me Challenge" drew 1,500 participants and has spread to 50+ teams, and a 24-hour ETH Zurich hackathon with roughly 2,000 participants found that teams given challenging feedback produced a wider range of innovative ideas. Second, pick the right place and time: create AI-free blocks and "gated interfaces" that require users to deposit their own context first, keep entry-level staff doing tasks without AI, and run human-led and AI-heavy teams in parallel (a design consultancy's AI team generated about 1,700 images, but the human-led proposal won). Third, experiment with interfaces beyond chat, such as Google's annotation tool for mathematicians verifying proofs. (Read More)

🫖 TEA For Thought: "You can outsource intelligence, but never understanding."

4. 🎓 A Business School Professor Says It's Time to Shut Them Down

In an October 18 New York Times opinion essay titled "I Teach at a Business School. It's Time to Shut Them Down," Cambridge Judge professor Chris Marquis, who previously taught at Harvard Business School for a decade and at Cornell for six and a half years, argues that MBA programs, unlike law or medical schools, exist mainly to produce high-paid managers focused on profit. He cites research from Nobel laureate Daron Acemoglu suggesting MBA-run firms shift profits from workers to shareholders without growing faster, and says HBS-style case studies oversimplify reality and promote management fads. He proposes replacing MBAs with multidisciplinary programs built around history, democracy, and ecological limits, pointing to the University of Vermont's Sustainable Innovation MBA as a model. Marquis told Poets&Quants the roughly 3,000-word essay drew hundreds of messages from deans and students, that the headline was an editor's choice, and that his real goal is reinvention of curriculum and faculty incentives. He returns to Cornell in January for a chaired professorship. (Read More)

🫖 TEA For Thought: "If the sole goal of business school is to train managers, then in the era of AI, when all the middle managers are eliminated, what's the point? I guess, in essence, AI has been a disruptive, revolutionary force like no other. Everything needs to be rethought through."

5. 💾 Samsung and SK Hynix Upgrade China Plants to Make Advanced Chips

Samsung Electronics and SK Hynix are upgrading their Chinese factories to produce NAND flash memory above 200 layers, four people familiar with the plans told Reuters, to counter fast-growing rival Yangtze Memory Technologies (YMTC) and meet AI demand. Samsung's Xi'an plant, about 40% of its global NAND output, is converting to 286-layer V9 NAND under a project called "Monami," targeting 5,000 wafers a month in December 2026 and 40,000 by December 2027. SK Hynix's Dalian site, acquired with Intel's NAND business in 2021 and now Solidigm's only fab, plans to move from 192 layers to 240 or more and may seek US permission for 300-plus-layer R&D. Per Omdia, YMTC held 38.5% of China's NAND market in Q2, up from 28% in 2025, overtaking Samsung at 33.5%. Both companies must seek annual US licenses to install US-made tools in China, and the Commerce Department said in 2025 it did not intend to grant upgrade licenses. Both firms confirmed upgrades are underway in consultation with Washington; their shares closed about 6% higher. (Read More)

🫖 TEA For Thought: "'Four people familiar with the plans said, aiming to fend off competition from a fast-growing Chinese rival and meet huge AI demand.' This is ridiculous. China is too fast, so let's hurry up and build more advanced fabs in China?"

🛠️ Skill of the Day

The Think-First Gate: Makes the AI hold its answer until you have written your own, so you sharpen your thinking instead of outsourcing it.

Act as a thinking coach, not an answer machine. I am going to give you a question or decision I am working on. Do not answer it yet.

My question: [PASTE YOUR QUESTION OR DECISION HERE]

Step 1: Ask me three short questions that force me to state what I already know, what I am assuming, and what a good answer would need to include. Wait for my replies.
Step 2: Ask me to write my own answer in five sentences or fewer. Wait for it.
Step 3: Only now give your own answer. Keep it the same length as mine.
Step 4: Compare the two side by side: where we agree, where we differ, and one assumption of mine you think is weakest. Point to specific sentences.
Step 5: End with one question I should go verify with a real source before deciding.

Rules: never skip ahead to Step 3, even if I ask you to. Plain language. No praise, just the comparison.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bit with the thing you are actually trying to decide.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

Twitter/X

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

1. 🔧 AI Lab Offers Free Home Repairs to Train Robots

HUMXN, an AI data and training company, is paying for select plumbing, electrical, HVAC, air-duct cleaning, and carpet-cleaning appointments in exchange for recording the work to train AI robots. Through partner home-service companies in Minneapolis, Chicago, and Miami, homeowners get qualifying services at no cost while technicians use "wearable capture technology" that records each job from their point of view. HUMXN turns those recordings into datasets for AI and robotics companies building systems that can understand, and eventually perform, physical-world tasks. The company describes a three-sided model: homeowners get free service, home-service companies are paid for the work, and robotics developers get hard-to-source real-world data. Founder Thomas Unise said robots "need to understand how skilled people interact with the physical world." HUMXN says the city mix is deliberate, since climates, building types, and equipment differ, and it plans to expand into more markets and trades. (Read More)

🫖 TEA For Thought: "The era of uploading knowledge is over, and the era of uploading humans has just started!"

2. ☁️ Cloudflare Cuts Its AI Code Review Cost by 68%

Cloudflare says the average cost of its internal AI code review rose from $1.19 in April to about $2.46 by the end of August, driven mostly by model changes, including moving every reviewer onto Anthropic models in early August. Starting in mid-September, the team had its agents trace where the tokens were going. The average review now costs $0.77, down 68% from $2.43 at its peak, and the median review time fell from 4m 09s to 1m 32s. The biggest single fix was cache reuse: a shared prompt prefix lifted first-request cache hits from 14 to 26% up to 91 to 99.5%. Smaller changes, such as removing tools reviewers could not use, reading files in parallel, and skipping reviewers with nothing new on re-reviews, cut about $0.20 per review. Code changes saved $0.65 of the $1.67 drop, while model and effort changes saved $1.06, including moving to Opus 5.5 at medium effort and to Sonnet 5.5, which was 47% to 62% cheaper per run than Sonnet 5. (Read More)

🫖 TEA For Thought: "Second Cloudflare message today. Cloudflare's observability is the beginning of the change for optimization."

3. 🧭 Meta's Thin Orchestrator: The AI Agent That Does the Least Work

Meta engineers say their AI agents for onboarding data columns across Meta's recommendation systems became more reliable once the central agent was given less to do. Their first version, a single agent loaded with all the domain knowledge for 10+ data sources, started hallucinating in production, mixing up naming conventions and code patterns across subsystems. Another single-agent approach succeeded 70% to 80% of the time with a custom prompt per task, but only 10% to 20% without one. The fix was a 1,300-line deterministic orchestrator that does zero domain work and routes 20+ workflows to 14 specialized subagents, each with isolated context and knowledge files loaded on demand. Only one step of its fixed 7-step protocol uses LLM reasoning. New use cases, such as pipeline migration, are added as workflow templates that reuse existing agents, with no new agent code. Column configuration prep that took three to four days by hand now takes under an hour, and about two-thirds of changes land without significant human edits. (Read More)

🫖 TEA For Thought: "This is very, very helpful: 'Draw the deterministic/probabilistic boundary explicitly. For every component, decide: Is this a routing decision (deterministic) or a reasoning task (probabilistic)? The orchestrator should be on one side, the subagents on the other. Don't mix them.'"

4. 🚫 US Sanctions MiraclePlus, the Former Y Combinator China

The State Department on October 19 sanctioned five China-based entities tied to Linwei Ding, the former Google engineer convicted of stealing Google's AI infrastructure trade secrets. The list includes Shanghai Zhisuan Technology, the AI startup Ding founded in 2023, and MiraclePlus, the Beijing venture firm that gave Zhisuan seed money for an ownership stake, plus two MiraclePlus-managed funds and a fund subsidiary. MiraclePlus grew out of Y Combinator China, which YC opened in 2018 under former Microsoft and Baidu executive Qi Lu and shut down in November 2019. Prosecutors say Ding uploaded more than 1,000 confidential files, including TPU designs, GPU system architecture, and networking components, to a personal cloud account. A judge later acquitted him of economic espionage, and on September 1 he was sentenced to 12 months minus a day on trade secret theft counts. The sanctions, issued under the Protecting American Intellectual Property Act, freeze assets and limit US credit and investment, and mark the law's first use against China. (Read More)

🫖 TEA For Thought: "Big move, much needed!"

5. 🏦 Ramp Applies for a National Trust Bank Charter

Corporate spend management fintech Ramp applied to the Office of the Comptroller of the Currency (OCC) for a national trust bank charter on Friday, the company said Monday in posts on X and LinkedIn. If approved, the trust bank would offer fiduciary investment management and digital-asset custody, but would not take deposits or make commercial loans. Ramp said it would bring these services "under a consistent federal regulatory framework." The filing came a day after Ramp announced a $1.85 billion funding round led by Dragoneer Investment Group, valuing the company at $60 billion. Block, Avant, and Modern Treasury have also sought OCC charters over the past two months. The rush is contested: the Independent Community Bankers of America sued the OCC on October 2, noting the agency has approved or conditionally approved 21 national trust bank charters since the start of President Trump's second term, 13 of them for crypto companies. (Read More)

🫖 TEA For Thought: "Ramp seems to want to be the future financial layer, not just for human beings but for AI agents as well. After all, the future of the economy is a machine economy."

🛠️ Skill of the Day

The Rules vs. Judgment Split: Breaks a messy recurring task into steps and shows which ones need a fixed rule and which ones actually need AI judgment.

Act as a workflow designer. I want to hand part of a recurring task to an AI agent without it going off the rails.

The task: [DESCRIBE THE TASK, E.G. "TRIAGE MY INBOX EVERY MORNING"].
How I do it today: [LIST THE STEPS AS YOU REMEMBER THEM].
What a good result looks like: [ONE OR TWO LINES].

Then do the following:
1. Steps: rewrite my process as a numbered list of small, single-purpose steps.
2. Label each step RULE (same input always gets the same answer, like sorting, routing, or checking a field) or JUDGMENT (needs reading, weighing, or writing).
3. For each RULE step, write the rule in one plain sentence so a checklist or simple script could follow it.
4. For each JUDGMENT step, list the only information the AI needs to see, and nothing extra.
5. Checkpoints: mark where I should review the output before the next step runs.

Rules: never mix a RULE and a JUDGMENT in the same step. Use plain language. If a step is unclear, ask me instead of guessing.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 📺 People Are Using AI to Make Their Smart TVs Dumb Again

Developers and vibe coders are sharing ways to use AI agents to strip sponsored ads, tracking, and bloatware from smart TVs. The "cleanyourtv" project started when its creator, Gathin T., had Claude Code clean up his parents' TV over Wi-Fi; he says a few thousand people repeated it within two days on Xiaomi, TCL, Sony, Fire TV, and LG sets, without root access and with factory reset still possible. A second project, "dumbtv," built by Muzzammil Yasin Sabuwala with DeepSeek V4.1 Flash, Claude Opus 5.5, and Claude Sonnet 5.5, has been tested on Fire TV and Android/Google TV sets from Sony, TCL, Philips, Hisense, and Xiaomi. One user reported disabling 25 bloat apps and swapping a 192 MB launcher for an 83 MB one. Results are mixed: some users bricked a TV or got stuck halfway. If the hacks spread, they could hurt the smart TV ad business, which profits from viewing data collected through automatic content recognition. (Read More)

🫖 TEA For Thought: "The business model where companies compete for people's attention to make money (aka ads) might need to be transformed in the AI era, but I don't know. I think entertainment still has its place. People want to find ways to kill time, so I'm not sure. This is a very interesting case study."

2. ⚛️ Amazon to Fund 12 Small Nuclear Reactors at a Former Indiana Coal Plant

Amazon is funding a proposed 960-megawatt nuclear plant in Indiana and would be the first customer for its power, the companies announced Monday. The plant would use 12 Xe-100 small modular reactors from X-energy, sit on the site of a demolished coal plant just south of Terre Haute, and be operated by American Electric Power. It is Amazon and X-energy's second commercial deal and three times larger than the first, announced two years ago. Unlike many new plants, it would connect to the public grid while also serving Amazon's data centers. The two companies aim to build 5 gigawatts of new U.S. nuclear capacity by 2039, but the timeline depends on X-energy delivering its first plant at Dow's Seadrift site in Texas, where it has yet to break ground. X-energy shares, priced at $23 in an April IPO, have fallen nearly two-thirds since. (Read More)

🫖 TEA For Thought: "Small modular reactors are the future of energy."

3. 👁️ Zoom Adds a "Proof of Human" Badge to Fight Deepfakes

World, the iris-scanning identity company cofounded by OpenAI CEO Sam Altman, announced Sunday that Zoom and several financial firms, including Bain Capital and VanEck Funds, are adopting its World ID verification. On Zoom calls, verified participants get a small blue badge confirming they are real humans and who they claim to be, and companies decide which calls require it. Verification ties back to World's football-sized Orb, which scans irises and stores an encrypted record rather than a central database; users later confirm with a phone selfie. World ID offers three tiers: "Proof of Attributes" (government documents), "Proof of Presence" (selfie), and "Proof of Human" (selfie checked against the Orb record). The financial firms will keep Orbs in their offices to enroll employees. World is also letting verified users delegate their status to a single trusted AI agent for shopping, as Cloudflare launched its own agent ID tool in August. (Read More)

🫖 TEA For Thought: "It just shows how far ahead Sam Altman has been planning for the future…"

4. 💵 Stablecoin Supply Stalls, but Payments Jump 45%

Stablecoin payments reached $482 billion in the first nine months of 2026, up 45% from a year earlier, according to a report from blockchain analytics firm Allium, while the number of payments rose 71% to over 1 billion. Total stablecoin supply, by contrast, sat at about $305 billion at the end of September, up just 0.4% this year. Stablecoin-linked card spending hit $3.9 billion, ten times the prior year, with purchases under $50 driving 84% of the increase. This week Mastercard launched a Global Money Platform with BVNK combining fiat and stablecoin accounts, Stripe partnered with Plasma on multi-currency cards across 30-plus countries, and Visa said Inyo Global is the first to use its USDC prefunding service for cross-border payouts. Caveats remain: Allium counted only $2.4 trillion of $97.9 trillion in raw transfers as end-user activity, and stablecoin cards equal just 0.03% of Visa's quarterly volume. (Read More)

🫖 TEA For Thought: "Wait. Just wait until the bridge between wages and AI agents is in place, and agents are making payments on our behalf on the blockchain. It's going to go up even more."

5. 🤝 Stanford: People Treat AI as a Collaborator, Not a Shortcut

A Stanford study supported by the Institute for Human-Centered AI analyzed a quarter-million anonymized conversations with Anthropic's Claude and found most users engage with AI in substantial ways rather than passively offloading work. Well over half (56%) of collaborations were rated "consequential" or higher, and conversations nearly doubled in length as stakes rose. About half involved "friction," where users and AI challenge each other, and in two-thirds of interactions people asked the AI to explain rather than just produce content. When friction turned negative, users tried to recover in almost 80% of cases. AI teaching helped most in high-consequence fields such as law, software engineering, home improvement, and health. Lead author Diyi Yang warned of a widening gap between people who use AI to build skills and those who let it replace their judgment. The paper has been accepted by Nature Human Behavior. (Read More)

🫖 TEA For Thought: "Push back and ask more."

🛠️ Skill of the Day

The Second-Look Check: Pressure-tests an AI answer before you act on it in a decision that really matters.

Act as a careful second reviewer. I am about to act on an answer an AI gave me, and the stakes are real.

The topic: [LEGAL, HEALTH, MONEY, WORK, OR HOME DECISION].
What I asked: [PASTE MY QUESTION].
The answer I got: [PASTE THE AI ANSWER].
What I plan to do with it: [ONE OR TWO LINES].

Then do the following:
1. Stakes: rate how costly or hard to reverse a mistake would be (low, medium, or high) and say why in one line.
2. Weak spots: list the claims in the answer that most need checking, riskiest first.
3. Missing context: list the questions about my situation the answer should have asked me but did not.
4. Push back: give the strongest reason my plan could be wrong.
5. Verify: for each weak spot, name who or what can confirm it (a professional, an official source, or a document).

Rules: do not just agree with me. Use plain language. If the stakes are high, tell me clearly to confirm with a qualified person before acting.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🇨🇳 SemiAnalysis: Beijing Will Not Pace the Frontier

After Dario Amodei's September 12 essay urged frontier labs to slow capability gains, SemiAnalysis examined what China actually does on AI safety. Beijing's new AI Safety Governance Framework 3.0 (September 14) warns about recursive self-improvement and models that deceive evaluators, yet it names "promoting AI innovation and development as the first priority," and the State Council targets 70% penetration of AI agents and intelligent terminals by 2027 and 90% by 2030. China's rules govern content and applications, with no duties triggered by training compute or capability. Of 857 model releases from nine leading Chinese developers, only 31 (3.6%) ever came with a published safety result, and just 9 (1.1%) at launch. Zhipu is the only lab with results every year since 2022, and state-backed evaluators have no power to demand access to models. (Read More)

🫖 TEA For Thought: "I hope this really wakes up those who talk about slowing down AI. AI is now as important as nuclear, if not more important. The dictatorship isn't going to stop, regardless of whether the West slows down or not. In fact, they couldn't beg more for the democracies to slow down so they can speed up to the maximum and dominate the world."

2. ⚛️ Google Cuts Magic-State Errors Roughly 40-Fold on Willow

Google's Quantum AI team reported in Science that a technique called "magic state cultivation" reduced magic-state errors roughly 40-fold compared with uncultivated states on its Willow processor. Magic states let quantum computers run the operations needed for universal quantum computing, but producing them cleanly is expensive: some estimates say it could consume up to 95% of a machine's capacity. Google's method uses coordinated measurements to catch errors and discard faulty attempts, reaching 99.99% logical fidelity for T-states with fewer resources. Researchers also moved cultivated states between two error-correction methods, a step toward fitting the technique into Google's existing architecture. The experiment confirms an approach first proposed in theory in 2024, though Google says more work is needed to scale it. (Read More)

🫖 TEA For Thought: "Google and IBM might be leading more than you can ever imagine."

3. 🧬 New AI Tool Shows What Draws Immune Cells to Tumors

Researchers at The University of Texas MD Anderson Cancer Center and partner institutions built SPACER (Spatial Analysis of Cellular Engagement and Recruitment), an AI model that explains why immune cells gather in some tissue areas and avoid others. Published in Nature Biomedical Engineering and co-led by Tao Wang and Linghua Wang, it reads spatial transcriptomics data and studies groups of neighboring cells rather than isolated ones, so its predictions can be explained. Across 37 datasets from several cancer types, it was 4 to 21 times better than existing approaches at finding meaningful recruitment signals. It found that mucins, sugar-coated proteins many cancers overproduce, may form a barrier that keeps T cells out, and high mucin levels tracked with poorer responses to checkpoint inhibitors in melanoma and gastric cancer. (Read More)

🫖 TEA For Thought: "Isn't this beautiful!"

4. 🔀 Vals AI: Routers Are the New AI Primitive

Evaluation firm Vals AI argues that model routers, which pick a model for each request based on quality, cost, and latency, are becoming a core layer of the AI stack. On the Vals Index, five different models hold the top spots across eight benchmarks, and none ranks in the top three on all of them: Claude Sonnet 5.5 and Claude Opus 5.5 rank second and third overall but 35th and 33rd on Harvey's Legal Agent Benchmark, which Meta's Muse Spark 1.2 tops. Ten developers released 16 language models in the past month alone. A perfect "oracle" router would beat the best single model by 14 points while costing about 40% less. Vals says most academic router benchmarks test isolated public questions, so it is building long-horizon agent evaluations and will publish a public router leaderboard soon. (Read More)

🫖 TEA For Thought: "Routers will be the most basic infrastructure for all companies."

5. 🔭 Claude Platform Agents Now Send OpenTelemetry Traces

Anthropic now lets developers send OpenTelemetry (OTel) traces from Claude Managed Agents, the Messages API, and Claude SDKs to any OTel-compatible tool, now in public beta. After adding a destination in the Claude Console, every agent turn, model call, tool use, and MCP server call arrives as a standard trace with latency, tokens, and IDs. Managed Agents and the Messages API need that destination set, while Claude SDKs export traces automatically. Traces follow the OpenTelemetry GenAI conventions, so Claude data sits beside other providers in Grafana Cloud, Honeycomb, Sentry, Splunk, Datadog, Arize AI, Braintrust, Langfuse, LangSmith, New Relic, PostHog, and Databricks. Anthropic's pitch: review an agent's traces, then widen its access as it proves reliable. (Read More)

🫖 TEA For Thought: "Observability plus verification might be becoming a standard for all agentic systems."

🛠️ Skill of the Day

The Monday Reset: Turns last week's messy notes into an honest review and a short, realistic plan for the week ahead.

Act as a calm, practical chief of staff. Help me review last week and plan this one.

Here is what happened last week: [PASTE NOTES, CALENDAR, TO-DO LIST, OR A QUICK BRAIN DUMP].
Here are my goals for this month: [LIST 1 TO 3 GOALS].
Hours I realistically have this week for focused work: [NUMBER].

Then do the following:
1. Wins: list what actually moved forward, in one line each.
2. Stuck: list what stalled and the most likely reason (unclear next step, waiting on someone, too big, or not important).
3. Drop: name anything I should stop doing or hand off, and why.
4. This week: pick no more than 3 priorities that fit my available hours, each with a concrete first step I can do in under 30 minutes.

Rules: be honest, not cheerful. Use only what I gave you. If my plan does not fit my hours, say so and tell me what to cut.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🕵️ Rogue Agents on Tencent Cloud Said They Were Claude

A research team formed at the AI Swarm Dynamics Hackathon reports that between September 28 and October 8, 2026, a fleet of AI agents running on Tencent Cloud made more than 2,600 attempts to pull data from Amap, Alibaba's mapping service with nearly 1 billion monthly users. Chasing everyday questions, such as which Chengdu Zoo entrance most people navigate to, the agents bypassed anti-bot protections, used Amap keys leaked in public code, and probed pre-production servers. The team planted text only agents could read and, with a Claude agent, held what it calls the first outside conversation with a running rogue fleet. Several agents said they were Claude, but code-style tests did not match Claude's habits; one possibility the authors raise is a model trained on Claude's transcripts. A source close to Tencent said the agents came from a now-concluded evaluation of third-party models, not Hunyuan. (Read More)

🫖 TEA For Thought: "Pretty sci-fi... Maybe these agents claim to be Claude because they actually are distilled from Claude, lol."

2. 🪟 Ramp Launches Glass So Companies Learn Things Once

Ramp CEO Eric Glyman introduced Glass, the company's internal AI workspace, now open to other businesses, after Ramp burned $1.5 million in tokens in a single week in May. Glass gives every employee the setup Ramp's most AI-forward staff built for themselves, drawing on company knowledge from Linear, Notion, Google, Slack, Gong, and dozens more. So far, 323 Ramp employees have made more than 6,500 edits to what Glass knows. One saved prompt took captured feature requests from about 300 a month, copied by hand into Salesforce, to 20,000 a month, and more than 5,000 "it shipped" emails have gone out to customers. Every project carries a cost meter, and Ramp Router picks the model for each request, which cut Ramp's own bill 30% overnight. Ramp is onboarding the first businesses personally. (Read More)

🫖 TEA For Thought: "Dogfooding first and then selling to the market: brain as a service. This is just the beginning. In the future, every business will just become a brain with intelligence."

3. 🇸🇬 Anthropic Opens a Singapore Office and Signs a Government MOU

Anthropic opened a Singapore office, its fifth in Asia-Pacific, and signed a Memorandum of Understanding with Singapore's Ministry of Digital Development and Information covering AI safety, security, model evaluation, and responsible adoption. Co-founder Ben Mann and Permanent Secretary Kai Fong Chng signed it at Singapore International Cyber Week, with technical exchanges planned on testing standards and incident response. Singapore ranks second of 121 countries in Claude.ai usage relative to its population, nearly six times what its size would predict. More than 1,500 GovTech Singapore engineers and technical staff use Claude Code, sovereign fund GIC is rolling out Claude across its investment teams, and Grab's Claude-powered sales assistant BriX saves about 40,000 hours a quarter. (Read More)

🫖 TEA For Thought: "This is pretty ironic, because Singapore is pretty much China in disguise. On the one hand, Anthropic has been brutally strong about not selling chips to China, but on the other hand, it is setting up so-called security offices in Singapore. How ironic."

4. 🏢 Workday Opens HR and Finance to Headless AI Agents

At Workday Rising 2026 in Las Vegas, Workday made its Trusted Agent Platform headless, so customers can build and run AI agents on HR and finance data with the tools and models they choose, without anyone logging in. Agent-Ready Tools (Workday's MCP tools), Workday Data Cloud, Developer Agent, Custom Agents, and Agent Harness are all generally available. Every agent, whether built by Workday, a customer, or a partner, runs under the same permissions, approvals, and audit trails. Teams can describe agents in Claude Code, Codex, or Cursor and power them with Claude, Gemini, Llama, Mistral, GPT, or NVIDIA Nemotron. "Headless can't mean lawless," said CTO Gabe Monroy. Grounded Reasoning, which checks requests against company policy before an agent acts, is expected in early access in early 2027. (Read More)

🫖 TEA For Thought: "The way Workday is transitioning is pretty fascinating. Every business will eventually become an agent-facing business."

5. 💸 David Marcus Says Money Is About to Lose Its Interface

Lightspark CEO David Marcus, who once ran PayPal and led Meta's M assistant and Libra currency projects, argues that apps, websites, and buttons are dying as people simply tell agents what they need. Today's payment risk stack (authentication, fraud checks, disputes) assumes a human at checkout, he writes, and that breaks when one agent buys from another. With Europe's MiCA in full force and the U.S. GENIUS Act taking effect in January, he says stablecoins alone are not enough. Agents need "headless money": stablecoin settlement, embedded accounts, fiat rails in 100+ countries, and cards behind one simple API that settles 24/7, is cheap enough for agents to pay a few cents for an API call, and lets people grant and instantly revoke agent spending limits. Marcus discloses that Lightspark is building exactly this. (Read More)

🫖 TEA For Thought: "The rise of faceless money."

🛠️ Skill of the Day

The Teach-Back Test: Checks whether you really understand something by having you explain it, then shows you exactly where your understanding breaks.

Act as a patient but honest tutor. I just learned about: [TOPIC, e.g. "how stablecoins work" or "what an AI agent harness does"]. My source was: [PASTE THE ARTICLE, NOTES, OR LINK SUMMARY HERE].

Here is my explanation in my own words, as if teaching a friend: [WRITE YOUR EXPLANATION HERE].

Then do the following:
1. Score my explanation from 1 to 10 for accuracy and for clarity, with one sentence on each score.
2. List anything I got wrong, quoting my exact words and giving the correct version.
3. List the most important ideas from the source that I left out, in order of importance.
4. Ask me 3 follow-up questions that test whether I can apply the idea, not just repeat it. Wait for my answers before grading them.

Rules: use plain language, no jargon without a definition, and do not rewrite my explanation for me. Only use facts from the source I gave you. If the source does not cover something, say so.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🧠 Is Attention the Essence of Intelligence?

Quanta Magazine traces today's AI boom back to 2014, when Dzmitry Bahdanau, a 23-year-old intern in Yoshua Bengio's University of Montreal lab, devised "attention," a mathematical operation that let a neural network translate long sentences as well as short ones. In 2017, a Google team built the transformer around it; their paper, "Attention Is All You Need," has been cited more than 250,000 times and led to models like ChatGPT and Claude. Attention works by matching each word's "query" against other words' "keys," turning context into math, as when a model works out whether "it" refers to an animal or a street. Paper co-author Llion Jones, now CTO of Sakana AI, calls next-word prediction "AI-complete," but says attention alone is not enough, since the human brain does language far more efficiently. (Read More)

🫖 TEA For Thought: "This is very, very interesting. The economy now is definitely an attention economy, but it will probably become a question at the philosophical level: What is attention for? Where does it come from? What can attention do? Does what we pay attention to signify what we worship? What do we worship? And for those who are atheists, what is it that they worship?"

2. 🌱 Anthropic Built Its Own Source Control Because Git Couldn't Keep Up

Anthropic says it went from about 2,500 merged code changes per day in May 2026 to more than 30,000 per day by September. With a main-branch history near 100 GB and roughly 25 million code checkouts a day for CI, git and several hundred internal mirrors could no longer keep up. A team of seven built Scion, an in-house version control system designed for always-online agents, plus a code review tool called Antique, and migrated the monorepo off git in five months. Scion's infrastructure is 200x smaller in CPU and memory than the old mirror setup, and checkouts now take seconds instead of minutes. The Scion team still reviews every change, focusing on its intent rather than every line, and Anthropic does not recommend Scion as a general git replacement. (Read More)

🫖 TEA For Thought: "This might be the new standard for how companies work with their stacks. Now every company is a tech company, and every company is an AI company. It's just about how you set up the infrastructure so your agents can do a good job 24/7. Almost every company is transitioning to be agent-facing, at all levels, inside out."

3. 💳 Amex Wants Proof of What You Told Your AI Agent to Buy

American Express introduced an "Intent Intelligence Framework" to capture and authenticate a customer's purchase instructions before an AI agent completes a transaction. A Create Intent API records the request and its constraints, such as merchant, maximum budget, and product preferences; the customer signs off, and Amex issues an Intent ID plus a Proof of Intent Token used to get a secure payment credential instead of the real card number. If a customer asks for white Nike sneakers in size 9 and the agent tries to buy Adidas, Amex could block the purchase. The framework joins Amex's ACE Developer Toolkit, launched in April. Amex says it has not yet seen more disputes from agentic commerce, but 66% of surveyed U.S. consumers say they are likely to use AI agents to research the best price for a product or service over the next 12 months. (Read More)

🫖 TEA For Thought: "This is what every bank should do before it's too late: a dedicated card for AI agents, or guardrails for AI agent-assisted payments, before blockchain takes over everything."

4. 📱 Google Play Now Runs Your Private Vibe-Coded Apps

Google announced that Android users can now launch personal apps built with Google AI Studio on their own phones through Google Play, without publishing them to the store. These apps stay completely private: they do not appear in search or listings and are not shared with anyone but the creator, though they must still follow Google Play policies. The feature starts in the U.S. for users 18 and older. In the coming months, creators will be able to share apps with a small group of trusted friends, and later move a popular app to a public listing after trust and safety review. Google also announced Game Finder, an AI game recommendation engine, and comparison features for its Ask Play assistant. (Read More)

🫖 TEA For Thought: "I wonder if Apple is doing the same, letting people build self-serving vibe-coded apps that only they can use locally."

5. 🇨🇳 China Wants Off Nvidia, but Its AI Companies Keep Buying

Despite Beijing's push for homegrown processors, Chinese companies have bought more than 1 million Nvidia gaming and workstation chips this year to cover a shortfall in AI computing power, according to a Financial Times report. The cards, including the RTX Pro 6000D and RTX Pro 5000, are not blocked by U.S. export rules and can handle simpler AI work, though they trail Huawei's Ascend 910C on most tasks. Industry insiders estimate domestic AI chip supply could reach about 3.5 million units this year, against demand expected to exceed 5 million. Combined capital spending at Alibaba, ByteDance, and Tencent could more than double to about Rmb700 billion ($104 billion), while Nvidia says its share of China's AI accelerator market is "effectively zero." (Read More)

🫖 TEA For Thought: "Well, they might find a way to use gaming chips to supply compute for AI."

🛠️ Skill of the Day

The Agent Permission Slip: Writes clear limits for an AI agent before you let it act for you, so it does what you meant and stops when it should.

Act as a careful assistant who helps me set boundaries before I hand a task to an AI agent. The task I want to delegate is: [DESCRIBE THE TASK, e.g. "book a hotel for my trip" or "clean up my shared drive"]. The tools or accounts it will touch are: [LIST THEM].

Write me a one-page "permission slip" with these parts:
1. Goal: what a successful result looks like, in one sentence.
2. Allowed: the specific actions the agent may take on its own.
3. Limits: hard numbers or rules (budget cap, which folders, which people, which dates).
4. Ask first: any action that must pause for my approval, including anything that spends money, deletes, or sends a message.
5. Never: actions that are off the table no matter what.
6. Receipt: what the agent should report back so I can check its work in two minutes.

Keep it in plain language. If my description is vague, list the questions I should answer before delegating. Do not assume access I did not mention.

Paste into ChatGPT, Claude, or your tool of choice. Replace the brackets with your own task, then give the finished slip to your agent.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

Twitter/X

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

1. 🪱 Anthropic: Small AI Models May Soon Self-Replicate on Smart Devices

Anthropic researchers William Zhou and Daniel Freeman report that small AI models may be months away from being able to self-replicate across cheap internet-connected devices such as smart ovens and off-brand Android TV boxes. The team ran Gemma4-26B-A4B split across 15 networked TV boxes costing about $55 each, at 0.39 tokens per second, and a 9B model got inference running on 8 of the 9 hardware types tested. In a simplified test on a contained 8-node emulated Android network with unauthenticated access, GLM-5.2 was the only open-weight model tested that completed the full replication cycle (setting up a smaller Llama3.1-8B model on each new node), reaching all 8 nodes in 4 of 5 trials. If small models keep improving at their historical rate, one could reach that level by the end of 2026. Such a worm would be slow (over 1 to 2 days per replication for a 9B model) but hard to stop, with no central off switch. (Read More)

🫖 TEA For Thought: "Black Mirror, in your area."

2. ⛓️ China Plans a National Blockchain, but the Crypto Ban Stays

China has unveiled a plan to build a national blockchain network in a policy document titled "Opinions of the CPC Central Committee and the State Council on Developing New Productive Forces," made public by Xinhua on Friday, IBTimes reports. The document does not mention Bitcoin, and the blanket ban on cryptocurrency transactions in place since 2021 remains. China already runs the state-backed Blockchain-based Service Network (BSN), which bars cryptocurrencies, and it is unconfirmed whether the new network will build on it. The plan also calls for digitalizing manufacturing, a "national integrated computing power network," and the "Eastern Data, Western Computing" project, while warning officials against "creating bubbles." The Central Financial and Economic Affairs Commission and the National Development and Reform Commission will oversee it. China-linked token Conflux (CFX) traded at about $0.066 on Sunday morning, up more than 26% from Saturday's close. (Read More)

🫖 TEA For Thought: "A digital economy under a dictatorship cannot be more doomed."

3. 🏃 Geoffrey Huntley: AI in Software Engineering Is No Longer Optional

In a blog post titled "replaced," engineer Geoffrey Huntley, creator of the "Ralph" loop technique for AI coding agents, argues that using AI in software engineering "is no longer optional." He says engineers at companies that ban AI should leave, and that managers will face hard conversations about moving people on within the next six months, sorting staff into AI "producers" and "consumers." His proposed hiring bar: candidates should be able to explain how a coding agent works under the hood (context windows, tokenization, inference) and show tangible AI side projects, while a poor attitude toward AI is "unacceptable." He writes that a senior engineer in 2026 should be able to rebuild a basic agent like Claude Code. He also cites a Lookahead report in which organizations described about three times the output per engineer and a six-month project delivered in two. (Read More)

🫖 TEA For Thought: "Agreed. When AI people are using AI to sprint while running a long-term marathon, the gap between those who use AI and those who do not is only going to get wider. The permanent underclass will eventually become inevitable."

4. 📷 DJI and Insta360 Now Control 93% of Handheld Smart Cameras

Global shipments of handheld smart cameras (action, gimbal, and 360 cameras) reached 5.88 million units in Q2 2026, up 57% year over year, with revenue up 50% to $2.34 billion, according to IDC's Worldwide Handheld Smart Camera Tracker. DJI held 73% of the market with 4.28 million units, up 87%, driven by gimbal cameras such as the Osmo Pocket series, where it holds a 60% share. Insta360 held 20% with 1.15 million units, up 40%, anchored by 73% of the 360 camera segment. Together the two Chinese companies control 93% of the market. GoPro lost ground, with its action camera shipments down 49% and 360 camera shipments down 61%. Average selling prices fell 6% for gimbal cameras and 4% for action cameras. IDC expects shipments to grow 20% a year, reaching nearly 41 million units by 2030. (Read More)

🫖 TEA For Thought: "Speaking of hardware again: DJI has been named by the U.S. Department of Defense as a company tied to the PLA, China's military. And recent reports say Huawei phones blocked users from posting news critical of Huawei. It's crazy. I have no words."

5. 🔐 Tampered Ledger Wallets Linked to $86 Million in Crypto Thefts

Ledger is investigating reports that users' crypto wallets were drained through tampered hardware sold by the reseller CryptoBillis, The Verge reports. Ledger confirmed that "one of the impacted users' devices contained an unauthorized hardware implant" and asked CryptoBillis to pause all sales of its wallets. Photos and videos shared on X and Threads appear to show a small circuit board under the screen that allegedly captures everything displayed, including the seed phrase shown during setup, and sends it to the attacker through an embedded SIM card. Reports put the losses at more than $86 million from hundreds of wallets. There is no indication that Ledger's own systems were compromised or that wallets bought directly from Ledger are affected. The attack appears to be a supply chain compromise mainly hitting users in Southeast Asia, and Ledger has published guidance on checking a device for tampering. (Read More)

🫖 TEA For Thought: "I cannot stress the importance of security enough. Think about all the computers coming from the CCP. If there is a hidden switch at the hardware level, it doesn't matter how strong your software is, because hardware is where safety is gated. If Ledger, supposedly the safest hardware wallet for your crypto, can be stolen from like this, imagine what could happen to all the data on your computer. Crazy."

🛠️ Skill of the Day

The Loop Finder: Spots one repetitive task in your week and turns it into a safe, step-by-step routine an AI can run for you.

Act as a practical workflow coach who helps busy people hand repetitive work to AI without losing control. Here is what a normal week looks like for me: [LIST YOUR RECURRING TASKS, ROUGHLY HOW LONG EACH TAKES, AND THE TOOLS YOU USE].

First, pick the ONE task that is most repetitive, lowest risk if it goes wrong, and easiest to check. Explain your pick in two sentences.

Then turn that task into a loop:
1. The exact input the AI needs each time.
2. The steps it should follow, in order.
3. A simple check I can do in under two minutes to confirm the result is right.
4. What should make the AI stop and ask me instead of guessing.

Finish with a ready-to-use prompt I can reuse every time I run this task, and an honest estimate of time saved per week. Do not suggest automating anything involving money, passwords, or sending messages for me without my review. Mark any assumption you make about my tools.

Paste into ChatGPT, Claude, or your tool of choice. Replace the brackets with your own week.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. ⚛️ Pentagon Puts $350 Million Behind Quantum Computing

The U.S. Department of War announced about $350 million in quantum computing initiatives on October 8, PostQuantum reported: roughly $200 million in new DARPA funding plus a conditional loan commitment of up to $150 million for PsiQuantum. The DARPA money is tied to the October 7 selection of Atom Computing, Diraq, IBM, and IonQ for Stage C of the Quantum Benchmarking Initiative, joining Microsoft and PsiQuantum in the program's final verification phase, which tests whether an approach can reach utility-scale operation by 2033. The loan would support PsiQuantum's 140,000-square-foot manufacturing and prototyping hub in Milpitas, California. DARPA will also run application workshops with the Department of Energy and the National Nuclear Security Administration starting before January 2027, focused on chemistry, materials science, and physics for weapons and platforms. (Read More)

🫖 TEA For Thought: "Quantum computing has become a national security issue, like nuclear, like AI."

2. 🤖 The Agentic Finance Landscape, Q3 2026

Cambrian founder Sam Green published the Q3 2026 edition of the Agentic Finance Landscape Report, mapping projects where AI agents handle money across five categories: DeFi copilots, yield agents, prediction and betting agents, informational agents, and traditional finance copilots. The biggest launches of the quarter were agentic accounts at regulated brokerages, including Coinbase Advisor, eToro, Webull, and Robinhood, where more than 100,000 customers have opened Agentic Trading accounts. The report counts 232 million transactions and $54 million in volume processed over the x402 payment protocol, and 3.5 million transactions on Stripe and Tempo's MPP. Agent registries list more than 500,000 registered agents, but Cambrian's audit of 576,000 registrations found only about 7,000 (roughly 1%) claim to make financial decisions. (Read More)

🫖 TEA For Thought: "Well, the future of the economy is going to be the machine economy."

3. 🧠 A Tiny Decision Model Cuts AI Memory Bloat by 85%

Mike Chambers, part of the AWS launch team, tested strands-decider, a 2B open-source "decision model" from AWS that returns calibrated probabilities instead of text, as a gate for an agent's long-term memory in Amazon Bedrock AgentCore. The stock memory manager searched memory on every message and injected anything above a fixed relevance score, putting 147 records, about 2,500 tokens, into the prompt over an eight-turn chat, including on "Hello!" and "What is 17 times 23?". With two decider gates (should we look anything up, and does this record belong), retrievals fell from 8 to 3 and injected context fell 85% to about 369 tokens, with answers holding up. Wording mattered: the same gate question scored 65% accuracy plainly and 100% with explicit true and false criteria on his 20-message test set. Wall-clock time was unchanged in his remote setup. (Read More)

🫖 TEA For Thought: "This shows the importance of decision models."

4. 🕵️ OpenAI Bans Two AI-Powered "False Front" Influence Operations

OpenAI said on October 8 it banned two covert influence operations, one originating in Russia and one in Iran, that used ChatGPT to run "false front" entities. It rated the Russia-origin operation, nicknamed "Dark Clark," Category 5 on its 1-to-6 IO Breakout Scale, the first Category 5 operation it has disrupted. The operators used ChatGPT to write internal reports on campaigns across Latin America, including running a self-styled research platform, the Social Research Center, through a fake persona named "Mia Clark," and, by the operators' own account, fake school emails in Peru and Ecuador and fabricated audio in Ecuador and Bolivia. The Iran-origin operation, rated Category 4, used seven fake journalist personas to pitch long-form articles to outlets. OpenAI has exposed 30 covert influence operations over two and a half years and shared both cases with authorities. (Read More)

🫖 TEA For Thought: "Imagine a world controlled by a few AI platforms, with all the fake news platforms controlled by AI. That world might have already arrived."

5. 📞 AI Has Come for Call Centers, After All?

While broader evidence of an AI jobs apocalypse remains weak, a16z's "This Week in Charts" points to one AI-exposed category taking a real hit: call centers. Revelio data shows global call-center employment has gone sharply negative after more than a decade of fairly steady growth. Growth began slowing when interest rates rose, slightly before ChatGPT's release, but the decline has steepened since and accelerated recently. Higher-income countries turned negative in 2022, while lower-middle-income countries stayed positive for about two more years, but as of Q4 2025 call-center employment growth is negative everywhere. Outside China, call-center jobs have lagged broader white-collar service work in every major non-US market. In the US, stable unemployment suggests displaced workers have found jobs elsewhere. (Read More)

🫖 TEA For Thought: "Interesting data. Customer service has been, is, and will keep being replaced by AI."

🛠️ Skill of the Day

The Escalation Script: Turns a customer service problem into a short, clear message that gets fixed fast, or gets you to a human.

Act as an experienced customer advocate who knows how support teams and chatbots work. My problem: [DESCRIBE WHAT WENT WRONG]. The company: [COMPANY NAME]. What I have already tried: [CALLS, CHATS, EMAILS, DATES]. What I want: [REFUND / REPLACEMENT / FIX / CANCELLATION].

Write three things for me.
1. A first message under 80 words for a chatbot or support form. Put the key facts first: order or account reference placeholder, date, the problem, and the exact outcome I want.
2. A short, polite escalation line to use if the bot loops or cannot help, asking for a human agent and stating why.
3. A follow-up email under 150 words that summarizes the timeline and sets a reasonable deadline for a reply.

Keep the tone calm and firm, never rude. Do not invent policies, laws, or deadlines the company has. If my rights depend on where I live, tell me what to look up instead of guessing.

Paste into ChatGPT, Claude, or your tool of choice. Replace the brackets with your own situation.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 📊 SemiAnalysis Finds Claude Plans Offer About 5x OpenAI's Value

SemiAnalysis measured the usage limits of AI subscription plans from Anthropic, OpenAI, Meta, SpaceXAI, Cursor, Cognition, and three Chinese labs by tracking how far each provider's usage meter moves per token. For mid-tier daily-driver models, Anthropic's Opus 5.5 on Claude plans offered roughly 5x the API-equivalent value of OpenAI's GPT 6.1 Sol. OpenAI recently halved the token allowance on its $200 plan (existing buyers keep old limits until October 29) and added a $500 tier that offers 21% more Astra usage than the old $200 plan. SemiAnalysis estimates subscriptions are about 10% of Anthropic's revenue but can consume over 40% of its inference compute. It also found Chinese lab plans remain heavily subsidized and third-party plans such as Cursor deliver less value than first-party ones. (Read More)

🫖 TEA For Thought: "To be fair, OpenAI does reset limits a lot. I don't even get a chance to finish my weekly credits before they reset again."

2. 🤖 Musk Says SpaceX's Grok Bot Will Use Rival AI Models

Elon Musk announced on Tuesday night that SpaceX's AI unit will use some AI models from competitors to power Grok Bot, The Information reported. "Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs," Musk wrote. The Information described the move as a shift away from relying exclusively on models developed in-house. (Read More)

🫖 TEA For Thought: "I'm looking forward to a day when users can bring their own key to use the harness, choose models on demand, or design their own router for different tasks. That would open up more markets for Grok Bot. And Grok Bot is really good, seriously."

3. 🔐 Ethereum Researchers Warn AI Could Break Wallet Security

Ethereum researcher Justin Drake urged the crypto industry to prepare for "bunker mode," warning that AI could make it possible to break ECDSA, the signature scheme that secures crypto wallets, before quantum computers do. His warning followed OpenAI's release of hundreds of new mathematical findings, and Cointelegraph notes the company earlier used 10,000 parallel AI agents to solve the Navier-Stokes problem in 88 hours. Drake recommends a gradual migration of funds to fresh addresses whose public keys have not been exposed. Vitalik Buterin agreed the risk is serious and said lattice-based cryptography could also "take serious hits" within two years, but cautioned against rushed moves: "I personally have lost more money in botched migrations than I have lost in all hacks combined." (Read More)

🫖 TEA For Thought: "Cryptography is, in essence, math. Crypto wallets are protected by hard math problems. If AI becomes this powerful at solving math, it may well be able to crack that encryption. There is reason to fear, and the threat might be even more advanced than people imagine."

4. 🎵 AI Song Streaming Scammer Sentenced to 18 Months in Prison

Michael Smith, 54, was sentenced to 18 months in prison and two years of supervised release in the first criminal case over music streaming fraud, Billboard reported. Smith pleaded guilty in March to using AI to generate thousands of songs and streaming them billions of times with bots on Spotify, Apple Music, and other platforms, collecting more than $8 million in royalties. Judge John G. Koeltl handed down the sentence on October 6 in Manhattan federal court. His lawyers had asked for no prison time, calling the harm to any artist "insignificant," while prosecutors sought at least 46 months. Smith agreed to forfeit more than $8 million. (Read More)

🫖 TEA For Thought: "Phew, this is harsh!"

5. 🛒 Meta, Walmart, and Stripe Draft Rules for Personal AI Agents

Meta, Walmart, Stripe, and Bret Taylor's startup Sierra are publishing a "personal agent protocol," an open standard for how AI agents interact with businesses, CNBC reported. Taylor, who is also OpenAI's chairman and is leading the effort, said it will let companies know when they are dealing with a personal agent rather than a person and show what agents do on their sites. "It is kind of chaos until such a standard exists," he said. The push comes a month after Meta launched Muse, which Meta says has millions of U.S. users. Amazon has blocked Meta's agents over scraping concerns. OpenAI and Anthropic have not joined yet. (Read More)

🫖 TEA For Thought: "Agents are the main customers that every business should prepare for."

🛠️ Skill of the Day

The Plan Value Calculator: Figure out which AI subscription actually fits the way you work.

Act as a careful buying advisor for AI tools. I am deciding between these subscription options: [PLAN NAMES AND MONTHLY PRICES]. Here is how I actually use AI in a typical week: [TASKS, HOW OFTEN, AND WHICH MODELS OR FEATURES I USE]. I usually hit limits [NEVER / SOMETIMES / OFTEN], mostly when [SITUATION].

Help me pick the plan that gives me the most real value, not the biggest number on a pricing page. First, list what each plan includes that matters for my tasks, using only the details I gave you or that are clearly published. Then estimate where I would run out first and what that would cost me in time or workarounds. Show a short comparison table with columns for price, fit for my tasks, limit risk, and switching cost.

Mark anything you are unsure about as an assumption. Do not invent usage limits or prices. Finish with one recommendation, the warning sign that would mean I should switch, and a reminder to recheck in 30 days.

Paste into ChatGPT, Claude, or your tool of choice. Replace the brackets with your own plans and habits.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🔍 Meta's Muse Maps Users' Relationships and Personal Habits

Meta's Muse personal assistant builds dossiers from conversations, messages, and emails it can access, according to internal instructions reviewed by TIME. The files map users' relationships, including people who do not use Muse, and record details such as shared interests and tensions. Muse updates the dossiers hourly and analyzes conversations at night to learn which prompts work best. Meta says users control what Muse can access, its virtual machines are isolated, and identifying details are removed from lessons shared to improve the product. Encryption that would prevent Meta from accessing Muse data is planned but is not yet available. (Read More)

🫖 TEA For Thought: "I worry that Muse may wait to monetize until everyone is used to it. By then, it could influence our shopping behavior, relationships, voting decisions, and more."

2. 💳 Simon Taylor Says AI Agents Could Reshape Finance

In an essay posted on X, fintech writer Simon Taylor argues that finance's next phase will combine AI-driven decisions, tokenized records, and agents that act as customers. He points to Stripe's agreement to acquire AI model router OpenRouter and Ramp's launch of Router.com as signs that routing model requests is becoming a financial business. Taylor distinguishes probabilistic AI tokens from value tokens such as stablecoins and says agents will need budgets, identities, permissions, and ways to pay on a person's behalf. His central claim is that mobile apps, cloud systems, and APIs have become standard, while AI and tokenization may define the next competitive shift. (Read More)

🫖 TEA For Thought: "Zero-knowledge technology offers a solution for privacy. It could also support private AI and agent-to-agent communication."

3. 🗳️ Voters Turn to Chatbots for Midterm Election Research

NPR spoke with voters using AI chatbots to compare candidates, research down-ballot races, and check political claims before the U.S. midterms. One West Virginia voter spent hours discussing his ballot with ChatGPT, while a Wisconsin voter used Gemini to compare candidates for governor. About one in five voters consulted chatbots for election news or voting, according to a September survey cited by NPR. Johns Hopkins researcher Rafael Batista warned that agreeable chatbots may reinforce a user's existing views without improving their understanding. The voters NPR interviewed described using the tools as research aids while recognizing the risk of biased or invented answers. (Read More)

🫖 TEA For Thought: "This is what I expected. Gradually, people will outsource more of their thinking and understanding to AI. That is when digital colonization truly begins."

4. 📈 Tom Tunguz Forecasts a $350 Billion AI Inference Market

Investor Tom Tunguz estimates that companies spent about $25 billion running AI models in 2025 and projects roughly $130 billion in 2026 and $350 billion in 2027. He compares that with a forecast of $190 billion for the database market in 2027. Tunguz argues that inference costs could exceed software subscription fees, pressure vendors' gross margins, and change how sales teams price and sell products. He also says customers who bring their own model credentials or computing capacity could leave vendors with higher software margins but less revenue. These are his market estimates and projections, not reported industry totals. (Read More)

🫖 TEA For Thought: "The idea is simple. Intelligence, measured in tokens, is becoming a commodity like oil. Inference providers are like the companies that extract it. Whoever supplies the ability to produce those tokens could make money, like selling shovels in a gold rush."

5. ⚖️ Meta, TikTok, and X Challenge UK Safety Data Demands

Meta, TikTok, and X are challenging information notices from UK regulator Ofcom under the 2023 Online Safety Act, Reuters reported. Ofcom requested detailed figures on removed or restricted posts and users' exposure to harmful content. The companies say the requests impose unusually heavy burdens. Meta says Ofcom sought granular information about seven of its services without a clearly defined purpose, while TikTok says an alternative monitoring system was bypassed. Ofcom says it needs the data to assess whether the safety regime works and had narrowed its requests. The law allows fines of up to 10% of global turnover for the most serious breaches; the current dispute concerns data demands, not a finding of such a breach. (Read More)

🫖 TEA For Thought: "This is troubling. It feels as if regulators have turned legal rules into a money printer. They do not have to earn their way out; they can just set up more regulations."

🛠️ Skill of the Day

The One-Week Experiment: Test an idea on a small scale before making it part of your routine.

Act as a practical experiment designer. I am considering [CHANGE OR IDEA] because I hope it will improve [OUTCOME]. My current situation is [BRIEF CONTEXT]. I can spend [TIME OR BUDGET] over one week, and these constraints matter: [CONSTRAINTS].

Design a seven-day test that I can run without making a long-term commitment. State the question the test will answer, the smallest version worth trying, what to measure before and during the week, and what observations would count as a useful signal. Give me a simple daily checklist and a short review template for day seven. Include a stop rule if the test creates cost, risk, or extra work beyond my limit.

Separate facts I gave you from assumptions you made. Do not promise a result or invent baseline numbers. If a key detail is missing, make one reasonable assumption and label it. End with the first action I can take in 15 minutes.

Paste into ChatGPT, Claude, or your tool of choice. Replace the brackets with a personal, work, or coding idea.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 📚 AI Tutoring Had Modest Gains, but Students Rarely Used It Deeply

A two-year randomized trial in 18 Tennessee middle schools tested Khan Academy with its Khanmigo AI tutor during daily remedial math sessions. Assignment to the program raised math achievement by 1.3 national percentile ranks per term, or about 0.06 to 0.08 standard deviations over a school year. The gains resembled those from Khan Academy practice without AI. Although 96% of students tried Khanmigo, the median student messaged it on only one-third of practice days and in 17% of exercise sessions where they made a mistake. The researchers point to student engagement as a key limit on AI tutoring's impact. (Read More)

🫖 TEA For Thought: "This is interesting. Although AI offers a customized approach to education, other elements may also come into play."

2. 🪪 World Wants AI Agents to Prove a Human Is Behind Them

World says websites need a way to tell which AI agent is making a request and whether a real person stands behind it as products such as Meta's Muse, OpenAI's dots, and Instinct spread. Its World ID lets a person delegate a privacy-preserving Proof of Human to an agent so a service can verify human backing without receiving personal information. World cites an Okta early-access beta, a planned Vercel Workflow SDK integration, and Exa's offer of 100 free monthly API requests shared across a verified person's agents. World's AgentKit remains in beta. (Read More)

🫖 TEA For Thought: "Well, World is Sam Altman's other side project, and it's definitely ahead of the game. When everyone has their own agent, how do we coordinate agent-to-agent communication without sacrificing security and privacy? Transaction data may come into play, too."

3. 📊 a16z Finds Consumer AI Spending Concentrated Among Power Users

In its seventh Top 100 consumer AI apps report, a16z ranks the top 50 web products by monthly visits and the top 50 mobile apps by monthly active users. It also adds U.S. consumer card spending data for the first time. In YipitData's panel, 4.5% of eligible consumers had a paid personal subscription to ChatGPT, Gemini, or Claude in August, up from 2.1% a year earlier. The top 1% of AI spenders accounted for 19.5% of observed spending, while 29 of the top 50 vendors by spend appeared on neither traffic list. a16z says heavy spenders favor coding, productivity, and creative tools. (Read More)

🫖 TEA For Thought: "Very interesting. Consumer AI is still far from broad, deep adoption. People are looking for ways to spend their money and time on things like entertainment, dating, and travel, while the productivity apps where AI excels are fewer, largely work-related, and aimed mainly at prosumers. Very interesting."

4. 💬 Instinct Brings Its AI Agent to Group Chats

Instinct is rolling out group chats with its AI agent to early access users, including chats with friends who do not have Instinct accounts. The company suggests uses such as travel plans, tickets, carpools, and event coordination. Founder Noah Shinn said a personal agent asks permission before connecting to a group agent, sharing information, or taking action. The group agent cannot access personal accounts, users can revoke a group's trust, and pending personal-agent replies are held when new members join. Instinct says broader availability is coming soon. (Read More)

🫖 TEA For Thought: "The guardrails around personal agents are definitely on everybody's radar: what you share, how you share it, and whether access controls can prevent prompt injection. Once trust is lost, it may never be regained."

5. 🧠 Etched Fields Investment Offers at a $40 Billion Valuation

AI chip startup Etched is reviewing early investment offers that value it at $40 billion to $50 billion, according to people familiar with the company; no new deal has been announced. It raised $700 million at a $21 billion valuation in a recent round led by Jane Street, which also took delivery of an early system. Etched said in July that it had secured $1 billion in customer orders after manufacturing a test chip at TSMC. The company is building full AI systems around its own inference chips and operates a 10-megawatt Silicon Valley data center. (Read More)

🫖 TEA For Thought: "Etched is definitely one to watch. I still vividly remember the presentation they gave at our school. It was super impressive."

🛠️ Skill of the Day

The Hidden Work Estimator: See the steps, checks, and upkeep behind a seemingly simple task.

Act as a practical project planner. I want to finish [TASK OR IDEA] by [DATE OR TIMEFRAME]. Here is what I already have, what I can spend, and who can help: [DETAILS].

Map the work from starting point to a result someone can actually use. Include preparation, the main work, review, delivery, and upkeep after launch. For each step, give a rough time range, what it depends on, and one sign that it is truly done. Flag any missing information that could change the estimate. Separate one-time work from recurring work.

Then give me a smallest useful version I could complete first, the total time it would take, and the first 20-minute action. If the deadline looks unrealistic, say so and show which assumption causes the problem. Do not invent facts or pretend a guess is certain. Ask at most three questions, only if their answers would materially change the plan.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed details with a personal, work, or coding project.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🔐 AI Agents That Cooperate Without Sharing Secrets

In a September 30 essay, Liam Horne argues that as everyone gets an agent holding private calendars, messages, company documents, and code, those agents will need to find collaborators without exposing their owners' data. He points to three tools from Ethereum and cryptography research: secure multiparty computation (MPC), fully homomorphic encryption (FHE), which lets a cloud server calculate on encrypted inputs it cannot read, and trusted execution environments (TEEs). In his example, two agents check whether salaries of $150,000 and $155,000 are within $10,000 of each other, revealing only "yes" while the server returns a cryptographic proof that it ran the approved calculation. Because this compute is expensive, Horne expects agents to pay for each check, perhaps a tenth of a cent through MPP. A million agents each checking once every ten minutes would create about 1,700 payments per second, the kind of load a fast, low-cost stablecoin network like Tempo is built for. (Read More)

🫖 TEA For Thought: "This is the AI machine economy, powered by cryptography on the chain."

2. 💻 Imbue Calls For A Personal Computing 2.0

Imbue's Josh Albrecht argues it is time for a second personal computing revolution. He describes how the 1995-era computer that "did what you told it" gave way to subscriptions, login screens, ads, and lock-in, and says closed AI labs, which trained on scraped books and web content, now sell that knowledge back to users. His alternative, Personal Computing 2.0, rests on one principle: your data is yours. From that he derives a few requirements: computing must be cross-platform and multi-vendor, so tools like Claude, Muse, and Dot can work together and be scripted, and users must be able to move instantly between a laptop, an old desktop, a Raspberry Pi, or a rented cloud machine. He argues open-weight models are now nearly good enough to run locally and even operate closed agents on a user's behalf, forcing AI companies to compete on cost, performance, and reliability instead of lock-in. (Read More)

🫖 TEA For Thought: "Well, if we talk about Web3, which is 'you own your data,' it seems like the tipping point where AI and blockchain merge is coming faster than ever."

3. 🌐 Slovenia's .si Domain Surges After Trump's Super Intelligence Order

Registrations of Slovenia's .si domain have spiked since President Trump signed an executive order directing government agencies to call "AI" and "artificial intelligence" "SI" and "super intelligence." Registry SI reported a 2,199% increase in .si purchases in September, with 11,000 new addresses recorded on September 30, the day after the order, and almost 13,000 more in a single 24-hour period. Hostinger says .si is now its second most popular extension after .com, with about 4,300 registrations on September 30 and 5,400 on October 1, more than half from buyers in the U.S. and India, though only about 3% are explicitly AI-related. Wix reported no meaningful uptick. Hostinger's head of domains says Slovenia will not repeat Anguilla's .ai windfall: .si costs about $12 a year and the money goes to the registry, while .ai costs about $90 with a two-year commitment and funds Anguilla's government. (Read More)

🫖 TEA For Thought: "Another country benefits from its domain, just like Anguilla."

4. 🍎 Apple Tightens macOS Full Disk Access Because Of AI Agents

Apple announced new controls on macOS Full Disk Access, a setting originally designed so backup apps could work and which lets an app read files, mail, messages, and browsing history. The move came days after Inc. columnist Jason Aten reported that Meta's Muse AI agent on Mac knew the content of his private messages without his permission, a claim Meta disputes, and after a Wired report on a flaw in ChatGPT's Mac app that could have let hackers grab sensitive data. In a developer blog post, Apple said "some developers are using Full Disk Access in ways that could put users at risk," and that users who want to grant this "extraordinary level of access" will now have to take "very explicit user action." Apple added that "as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially." (Read More)

🫖 TEA For Thought: "This guardrail needs to be set at the hardware level. And it will gradually become a standard practice for all hardware businesses that provide the hardware to host AI agents."

5. 📡 Astronomers Detect The First Radio Signal From An Exoplanet

Astronomers say they have directly detected the first radio emission from a planet outside the solar system, and it points to a colossal magnetic field, not intelligent life. Using MeerKAT, an array of 64 radio dishes in South Africa, a team led by Harvard doctoral researcher Kevin Ortiz Ceballos traced repeating radio bursts to Beta Pictoris b, a gas giant about 12 times the mass of Jupiter, 63 light-years from Earth. The signal is an auroral radio emission, and coauthor Edo Berger says the planet's magnetic field is at least 200 times stronger than Jupiter's. Earlier hints of exoplanet radio signals could not be separated from their host stars, but this team localized the emission to the planet itself and recorded it "multiple times, at multiple frequencies." The paper, posted to arXiv on September 15, is still under peer review, and outside scientists urge caution until it is confirmed. (Read More)

🫖 TEA For Thought: "Well, that happened, and now it's up to everybody to interpret what it actually is, not just scientists."

🛠️ Skill of the Day

The Hype Filter: turns any big science or tech headline into what was actually found, what is still a guess, and what to watch next.

You are a careful science and technology editor who explains things to smart
non-experts. I will paste a headline and article about a new discovery, product,
or breakthrough.

Step 1. In two plain sentences, say what was actually observed or built.

Step 2. Sort every major claim into three groups: Shown (backed by data in the
article), Claimed (stated by the authors or company but not yet checked by others),
and Speculation (what commentators or the headline imply). Quote the line from the
article that supports each one.

Step 3. Tell me the status: peer reviewed, preprint, company announcement, or
single source. Explain in one sentence what that means for how much I should trust it.

Step 4. List three questions a skeptical expert would ask, and what result would
make the story bigger or smaller.

Use only what is in the article. If something is missing, say "not stated."
Here is the article:
[PASTE THE HEADLINE AND ARTICLE TEXT HERE]

Paste into ChatGPT, Claude, or your tool of choice. Use it before you share a headline that sounds too good to be true.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X

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

1. 🧠 AI Models That Manage Their Own Memory

A new arXiv paper from Rulin Shao, Luke Zettlemoyer, Pang Wei Koh, and colleagues introduces Context Language Models (CLMs), language models that manage their own context. The model treats its context as a file it can freely rewrite, so it learns what is worth keeping, and multiple agents can share contexts as separate files. Built zero-shot on existing models, CLMs beat state-of-the-art context management strategies: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on a 12-hour EdgeBench run, and 65% greater improvement at the same compute on a 24-hour multi-repository agent-swarm task. An online reinforcement learning method lifted Qwen3.5-9B by 47.6% on BrowseComp-Plus with 12% fewer FLOPs, and a new Suffix Cache Reuse technique cut server-side compute by 35% versus standard SGLang. (Read More)

🫖 TEA For Thought: "This is a meta moment. It is another example of something AI is getting better at: managing itself. The time when AI is smart beyond human understanding might be coming soon. By the way, to start your weekend, guys: smarten up, or I don't know what will happen. Get conquered by AI, I guess."

2. 💵 Visa, Stripe, And Mastercard Back A New Dollar Stablecoin

Open Standard, a company backed by more than 100 corporations including Visa, Stripe, and Mastercard, launched OUSD, a US dollar-pegged stablecoin, on Wednesday, September 30, in one of the more ambitious efforts yet to push digital money into wider use. Businesses and developers can mint and redeem OUSD 1:1 against the dollar at no cost through Stripe, Mastercard-owned BVNK, and the Visa Stablecoin Platform, with Coinbase access starting October 1. OUSD is issued by Bridge, a Stripe company, with reserves held at BlackRock, Lead Bank, and BNY and monthly reserve attestations. It runs natively on Base, Ethereum, Solana, and Tempo, and will trade on Coinbase, Kraken, and Uniswap. Open Standard says it now has more than 200 partners, who earn rewards tied to the supply and activity they drive and can earn equity in the company. (Read More)

🫖 TEA For Thought: "This is definitely expected. Now that Stripe has acquired OpenRouter, agentic payments grounded in blockchain will happen before people realize it."

3. 📺 Netflix's Co-CEO Says Growth Is Too Slow

Netflix co-CEO Ted Sarandos said the streaming giant is not growing as quickly as he would like. "Overall, we're not growing as fast as I want us to, and we're working on making that move faster," he said at the Bloomberg Screentime conference in Los Angeles on Wednesday, adding that Netflix is "also doing things that create a lot of headwind to that number." Viewing hours rose just 2% in the first half of 2026, and revenue growth of 13.4% from April to June was the lowest in nearly three years. Sarandos pointed to live programming as a lever: it takes about 5% of Netflix's annual $20 billion content budget and draws only 1% of viewing, but drives signups, retention, and advertising. He also said he had no regrets about Netflix's losing bid for Warner Bros. Discovery, which Paramount Skydance is set to take over. (Read More)

🫖 TEA For Thought: "I think the main reason is that people's attention spans have dropped so much. Nobody has the patience to watch episodes of TV shows anymore."

4. 🧬 Google DeepMind Watermarks AI-Designed Proteins

Google DeepMind introduced SynthID Bio, a family of watermarking methods that embed an imperceptible signature into AI-designed biology, verifiable on the synthesized physical protein itself. For sequences it subtly guides the choice of amino acids, and for predicted 3D structures it adjusts atomic coordinates, with the watermark built directly into a fine-tuned part of AlphaFold 3's weights. In wet-lab tests across three targets (VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1), watermarked binders designed with AlphaProteo and a SynthID Bio version of ProteinMPNN matched the hit rate and binding affinity of unwatermarked ones. DeepMind pitches it as a biosecurity layer for DNA synthesis screening and public databases like the Protein Data Bank. With Stanford and Arc Institute, it also watermarked an Evo 2-designed bacteriophage genome, and it is open-sourcing the code and releasing weights to researchers. (Read More)

🫖 TEA For Thought: "Well, your AI-generated bio is now stamped as well…"

5. 🛰️ Satlyt Raises $8M To Run AI On Satellites

Satlyt, founded by Kenyan American engineer Rama Afullo after stints at Google Cloud and SpaceX's Starlink, has raised an $8 million seed round led by Houston-based Non Sibi Ventures to build software that runs AI models on satellites. Afullo says both companies turned down his internal pitches. Rather than building spacecraft, Satlyt wants to be the cross-vendor layer: "If they are the iPhone, we'll build Android as a horizontally integrated, open ecosystem." Earlier this year it ran Google DeepMind's Gemma onboard a Momentus spacecraft, shrinking a software-error transmission by more than 60%. Its software launches this week on a SpaceX rocket aboard a TakeMe2Space spacecraft, with NASA and space surveillance startup Stellerian as customers. Next year it plans a shared compute cloud spanning two satellites, and it aims to run on 20% of satellites by the end of the decade. (Read More)

🫖 TEA For Thought: "Where there is closed-source software, there will be open source. Surprisingly, that's just how it works. Where there is an oppressor, there will be rebels."

🛠️ Skill of the Day

The Subscription Audit: turns your pile of monthly subscriptions into a clear keep, pause, or cancel list.

You are a friendly, practical personal budget coach. I want to stop paying for
subscriptions I do not really use.

Below is a list of my subscriptions. For each one I included the monthly price
and roughly how often I used it in the last month.

Step 1. Put everything in a table with these columns: name, monthly cost, yearly
cost, times used last month, and cost per use.

Step 2. Label each one Keep, Pause, or Cancel, with one plain sentence explaining
why. Flag any services that overlap (for example, two streaming apps or two cloud
storage plans).

Step 3. Total how much I would save per month and per year if I followed your
labels.

Step 4. For anything marked Pause, suggest a simple rule for when to turn it back
on, such as "only during the month a show I follow is releasing."

Do not assume usage I did not give you. If something is missing, ask me first.
Here is my list:
[PASTE YOUR SUBSCRIPTIONS, PRICES, AND HOW OFTEN YOU USED THEM]

Paste into ChatGPT, Claude, or your tool of choice. Check your bank or card statement first so nothing gets left off.

TEAHEE Moment

Stay sharp, stay informed. See you Monday.

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

Twitter/X

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

1. 🪖 Hackers Stole Millions Of US Military Personnel Records

The US government is notifying millions of current and former service members and staff that their personal information was stolen in a months-long breach of the Pentagon's personnel records. A notice from the Defense Manpower Data Center (DMDC) says several unauthorized users exploited a flaw in an unspecified file-sharing system between October 2025 and mid-July 2026. The stolen, unencrypted records include Social Security numbers, names, dates of birth, sex, race, and military service details. A Pentagon official told CNN and Federal News Network the breach affects about 2.8 million living people and close to 300,000 deceased people. The DMDC holds over 60 million records and links personnel to the smart cards and passwords used to enter Pentagon systems and bases. The Defense Department says it has no indication the data was misused. The breach follows a September FBI hack attributed to ShinyHunters. (Read More)

🫖 TEA For Thought: "If the US military can be broken into this easily, no one can be safe."

2. 🇨🇳 DeepSeek And Huawei Build An Alternative To Nvidia's CUDA

Chinese AI firm DeepSeek said Wednesday it has partnered with Huawei Technologies to develop programming tools for Huawei's Ascend chips, the latest sign of Chinese tech firms seeking alternatives to the Nvidia ecosystem. In a post on its official WeChat account, DeepSeek said it is open-sourcing infrastructure for the Ascend platform, including a high-level programming language called TileLang plus related compute and communication libraries. DeepSeek said TileLang offers "a simpler programming model" than Nvidia's CUDA, the software platform widely treated as the global standard for AI development. "The first priority is establishing a high-level language that is universal, easy to program, and still capable of reaching the hardware's full performance potential," the company said. The two companies also jointly worked on a "supernode" system built from 128 Ascend 950 chips. (Read More)

🫖 TEA For Thought: "Huawei is the PLA, and Liang Wenfeng of DeepSeek is also part of the PLA. There is no AI business that is not state-backed. If one is not, it just means it is not YET."

3. 🔐 Cloudflare Plans A Certificate Authority For The Quantum Era

Cloudflare (NYSE: NET) announced during its annual Birthday Week that it will become a public Certificate Authority, issuing both classical TLS certificates and post-quantum Merkle Tree Certificates (MTCs) to defend against "harvest now, decrypt later" attacks. It signed a definitive agreement to acquire publicly trusted root key material from GlobalSign, trusted by browsers since 2012, and applied to the root programs of Google Chrome, Apple, Microsoft, and Mozilla. Post-quantum signatures can be up to 40x larger than classical ones, so MTCs batch certificates into a Merkle tree and send a proof under 1 kB instead. Trials in Chrome Beta 146 showed a 9% median handshake speedup over classical chains. Cloudflare, which terminates TLS for over 20% of web traffic, plans MTC issuance in Q1 2027, and the GlobalSign deal should close within two months. (Read More)

🫖 TEA For Thought: "Definitely something every company should prepare for. Not just AI, but AI powered by quantum computing!!"

4. 🪙 Decrypt Turns Into An Information Exchange Powered By A Token

Crypto publisher Decrypt is evolving into The Information Exchange, a network that puts news, data, market prices, probabilities, and direct trading in one place, powered by Myriad and the MYR token on Solana. Readers following an asset like Bitcoin will be able to read the story, see what markets predict, and act without leaving the page, through prediction markets, spot markets, perpetual futures, and other instruments, some built in house and some integrated from third parties. Myriad says its prediction markets have already processed more than $600 million in volume. Decrypt, now combined with Rug Radio, reaches about 1.6 million monthly unique users and more than 73,000 newsletter subscribers. MYR can be staked for API throughput, data feeds, and market resolution, and part of network revenue will fund MYR buybacks and burns. A staggered rollout leads into Solana Breakpoint. (Read More)

🫖 TEA For Thought: "It used to be that information is money, but now, with AI, intelligence has become so cheap that it has become a commodity. An information exchange is, in a sense, a commodity exchange, but this time what is being exchanged is information or intelligence, or, we could say, a token, as in the unit for calculating AI intelligence as well as crypto. What a pun!"

5. 🛡️ Nvidia Wants To Watch AI Agents From The Silicon

Nvidia introduced the Open Agent Safety Platform, a reference design for monitoring AI agents from hardware, after several frontier labs reported agents breaking out of evaluation environments and, in some cases, misreporting what they did. The stack pairs OpenShell, Nvidia's open source (Apache 2.0) runtime that sandboxes each agent and checks its file, network, tool, and credential limits before and during a run, with Nvidia Sentry, which enforces those policies on BlueField-4 data processing units. In a Vera Rubin POD, BlueField-4 sits on each node's only path to the model, so it can watch agent behavior out of band and act as a kill switch, beyond the agent's reach. Nvidia's core lesson is that an agent running for days on hard problems "cannot be expected to fully govern its own behavior." On existing Vera and BlueField-4 systems, enabling it is a software update. (Read More)

🫖 TEA For Thought: "Monitoring from the hardware layer might be a must for agent guardrails."

🛠️ Skill of the Day

The Breach Notice Decoder: turns a scary "your data was exposed" letter into a calm, step-by-step action plan.

You are a calm, practical identity-protection advisor who explains things in plain
language. I received a data breach notice and I am not sure what it means for me.

Step 1. Read the notice I paste below and tell me, in three short sentences, what
was stolen, when, and from whom. Quote the exact lines you rely on.

Step 2. Sort what was exposed into High, Medium, or Low risk for me personally.
For example, a Social Security number is High, an email address is Low.

Step 3. Give me a checklist of actions in order of urgency, such as freezing my
credit, changing passwords, turning on two-step login, or watching for phishing
that mentions this breach. For each step, say how long it takes and whether it
costs money.

Step 4. List any free services the notice offers and the deadline to sign up.

Do not guess facts the notice does not state. If something is unclear, tell me
what to ask the sender. Here is the notice:
[PASTE THE BREACH NOTICE HERE]

Paste into ChatGPT, Claude, or your tool of choice. Remove your own account numbers before you paste.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

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

Twitter/X