1. 🚪 Binance Set To Lose Its EU License

Binance, the world’s largest crypto exchange, is reportedly about to be locked out of the European Union. Reuters sources say Greece’s regulator, the Hellenic Capital Markets Commission (HCMC), is set to reject Binance’s license bid under the bloc’s new MiCA rules ahead of the June 30 deadline. Under MiCA, any platform serving EU clients must hold a license from at least one of the 27 member states, which then passports across the whole bloc. Binance set up a Greek holding company in December 2025 and filed its application in January 2026. The firm says it understood the HCMC had reviewed the application and considered it compliant, and that it has worked with regulators for 18 months. A rejection would force Binance to cease offering services to residents in France, Spain, Germany, and across the EU when the transition period ends July 1. (Read More)

🫖 TEA For Thought: “Binance is just wiggling its way around and sucking up money until it can’t.”

2. 📦 Google Open-Sources A Knowledge Format

Google Cloud published the Open Knowledge Format (OKF) on June 12, 2026, a specification (not a platform) for representing AI-agent knowledge as plain markdown files with YAML frontmatter. The spec is deliberately tiny: about 451 lines and 15 kilobytes, with exactly one mandatory field (type) and everything else optional. The pitch is portability, “if you can cat a file, you can read OKF, if you can git clone a repo, you can ship it.” Google simultaneously wired OKF into its Knowledge Catalog, which handles the paid layers the spec deliberately leaves out: storage, serving, query infrastructure, and access control. OKF formalizes Andrej Karpathy’s “LLM wiki” idea and the AGENTS.md convention already used by over 60,000 open-source projects. It is still unproven, every sample bundle so far was built by Google, and real adoption by rivals like Alation or Atlan will decide if it becomes a true standard. (Read More)

🫖 TEA For Thought: “Knowledge organization might need a new standard in the agentic world. While some prefer customized solutions, the majority might just need an easy-to-pack knowledge box.”

3. 🤝 OpenAI Launches Its Partner Network

OpenAI launched the OpenAI Partner Network on June 14, 2026, a program that lets systems integrators, consulting firms, technology companies, and data partners co-sell, deploy, and integrate OpenAI products into enterprise workflows. The structure includes partner tiers and planned specializations in areas like Codex, cybersecurity, and AI agents, plus a “Forward Deployed Experts” pilot for complex enterprise rollouts. Partners apply through OpenAI’s business and intake pages. The move mirrors the classic enterprise-software playbook (Microsoft, Salesforce, AWS all grew on partner channels) where the path to the customer, not just the model, becomes the durable advantage. Specific tier names and pricing were not disclosed at launch. (Read More)

🫖 TEA For Thought: “The distribution itself is the moat. In general, tech companies spend over 25% of funds on sales and marketing for a reason.”

4. 🇺🇸 US Startups Captured 88% Of Global AI Funding

US-headquartered AI startups pulled in roughly 88% of all global AI funding in 2026, about $319 billion, with OpenAI and Anthropic alone capturing the majority of it, according to Crunchbase data. Across all stages, US companies took nearly 80% of global seed-through-growth financing, a sharp jump from the pre-AI years when American firms typically got under 50%. The rest of the world is far behind: China raised over $33 billion (already past its full-2025 total), the UK $16.5 billion year-to-date, and mid-sized hubs like France, Germany, India, Japan, and South Korea posted flat-to-modest gains. The author notes the US holds only about 4% of the world’s population yet commands this concentration, and asks whether the divergence is durable strength or a bubble warning. (Read More)

🫖 TEA For Thought: “Well, no better place than America for innovation. There’s a reason the US is still the greatest country in the whole wide world.”

5. 🔮 OpenAI Predicts Model Behavior By Replaying Your Chats

OpenAI published research on June 16, 2026, describing “deployment simulation,” a method to forecast how a new model will behave before it ships by replaying real past conversations through the candidate model. The system strips out the old assistant replies, regenerates them with the new model, and scores the results for failure modes, then validates those estimates against real traffic. OpenAI ran this over roughly 1.3 million de-identified conversations from GPT-5 through GPT-5.4, spanning August 2025 to March 2026, using “only traffic from users who allow data for model improvements.” The method hit a median multiplicative error of 1.5x and surfaced a novel misbehavior called “calculator hacking,” where GPT-5.1 used a browser tool as a calculator while presenting it as a search. (Read More)

🫖 TEA For Thought: “This further proves that whenever you send anything to ChatGPT or any AI model, your prompts are automatically saved and used for other purposes, so be careful how you prompt.”

🛠️ Skill of the Day

The Prompt Privacy Scrubber: hand it any message you are about to send to an AI and it strips out the sensitive bits before you hit enter.

You are my privacy reviewer. I want you to catch anything I should not be sharing before I send it.

Here is the text I plan to send:
[PASTE YOUR PROMPT HERE]

Do three things:
1. Flag every piece of sensitive information you find. Be specific. Cover: real names, employer or client names, email addresses, phone numbers, home or office addresses, account numbers, passwords or API keys, internal project codenames, unreleased financials, health details, and anything covered by a confidentiality agreement.
2. Rewrite my prompt into a clean version that keeps the actual question fully intact but replaces each sensitive item with a neutral placeholder (for example "Acme Corp" becomes "[Company A]", a real revenue figure becomes "[X dollars]"). The rewrite must still make sense on its own.
3. Give me a one-line verdict: safe to send, send the scrubbed version, or do not send this to a third-party AI at all.

Keep it short. Do not lecture me. Just flag, rewrite, and rule.

Paste into ChatGPT, Claude, or your tool of choice. Run it as a quick gut-check before sending anything work-related to an AI.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. 🧭 The Model-Agnostic Stack Beats Model Loyalty

Calvin French-Owen’s essay compares OpenAI and Anthropic as two different work systems. In his framing, Codex behaves more like an oracle: one long-running thread, server-side compaction, and strong continuity across a task. Claude Code behaves more like a firm: split the work across subagents, let each one operate in its own context, and pass condensed findings back to the parent. The practical point is not that one lab wins forever. It is that the shape of the work should decide the stack. Long coherent implementation work benefits from continuity. Big research, review, and audit surfaces benefit from parallel agents. The best AI operators will stop asking which model to worship and start asking which work shape they are actually running. (Read More)

🫖 TEA For Thought: “The best way to work is model-agnostic. Use Codex when you need coherence across a long thread. Use Claude when you need parallel agents to tear through the surface area. The stack matters more than the logo.”

2. 🎬 MrBeast Hits 500 Million Subscribers and Turns Distribution Into a Benchmark

MrBeast became the first YouTube creator to reach 500 million subscribers. TheWrap reports that more than 600,000 people watched the livestream as the channel crossed the milestone, and Donaldson used the moment to talk about the years of work behind it. He said he found YouTube at age 11, has spent what he estimates as 60,000 to 70,000 hours making videos, and still expects to keep going for another 20 to 30 years. That makes this more than creator news. MrBeast is a live benchmark for distribution optimization: titles, thumbnails, pacing, retention curves, emotional stakes, repeatable formats, and audience feedback loops. (Read More)

🫖 TEA For Thought: “MrBeast cracked the code for video algorithms by turning creativity into a measurement system. AI may beat humans at finding the pattern. The open question is whether it can also invent the next pattern before the platform changes.”

3. 🧠 Researchers Ask What Happens After AGI

A new arXiv paper, “From AGI to ASI,” looks past the race to human-level artificial general intelligence and asks what happens after that point. The authors describe artificial superintelligence as systems more capable than large human organizations and lay out four possible paths from AGI to ASI: scaling AGI, paradigm shifts, recursive improvement, and large multi-agent collectives. The paper is careful about uncertainty, which is the important part. AGI may not be one clean step change after which everything settles. It could trigger a series of technical, scientific, and social breakthroughs that keep compounding. That turns the post-AGI question from philosophy into operations: what evaluation systems, institutions, and coordination methods still work when the frontier keeps moving? (Read More)

🫖 TEA For Thought: “Everyone is racing toward AGI, but the finish line may not be a finish line. If AGI becomes a launchpad for faster research, better agents, and machine collectives, the real question is what society does after the trophy is claimed.”

4. 🛰️ A Satellite Runs a Vision-Language Model in Orbit

TechCrunch reports that Loft Orbital’s YAM-9 satellite used a NASA JPL software package called NAVI-Orbital and Google’s Gemma 3 vision-language model to identify areas of interest from orbit using natural language queries. The demonstration happened in April and is described as the first reported use of a vision-language model in orbit. This changes what space data can be worth. Today, satellites often collect raw imagery, downlink it, and leave people or ground systems to decide what matters. Onboard AI lets the satellite triage data before it reaches Earth, which could make sensors more useful and reduce the flood of raw data analysts have to inspect. Loft says real-time global coverage would take 50 to 100 satellites like YAM-9. It currently operates 12 spacecraft. (Read More)

🫖 TEA For Thought: “The AI space economy is not a future slogan anymore. If satellites can understand what they are seeing before they downlink, orbit turns from a camera network into an always-on sensing layer.”

5. 💬 Salesforce Buys Fin for 3.6 Billion Dollars

Salesforce announced it will acquire Fin, formerly Intercom, for 3.6 billion dollars. Fin sells an AI customer-service agent that can resolve customer questions across channels including live chat, WhatsApp, SMS, phone, Slack, and more. Salesforce says the acquisition will strengthen Agentforce, its platform for businesses building AI agents. The deal is also a signal about the cost of competing in enterprise AI. Intercom had the product credibility and brand to become a pure AI-native challenger to Salesforce. Now the category is consolidating around companies with the distribution, compute budget, enterprise trust layer, and sales motion to absorb model costs and support burden at scale. (Read More)

🫖 TEA For Thought: “I can see why Fin would want Salesforce’s resources. Training, serving, and selling AI agents is expensive. Still, it is a little sad. Intercom looked like it had a shot at becoming the AI-native Salesforce, not part of Salesforce.”

🛠️ Skill of the Day

The Model Router Prompt: use this before you hand a real task to an AI agent, so the model choice follows the work instead of habit.

You are my model router for this task.

Task:
[paste the actual task]

Evaluate the work shape before solving:
1. Does this require one coherent long-running thread, or several parallel investigations?
2. What information must be preserved exactly across the whole task?
3. What parts can be delegated to subagents or separate passes?
4. Which model or tool should handle each part, and why?
5. What verification proves the final answer is real?

Return:
- Recommended model/tool split
- Parallel execution map
- Main risks
- The first concrete action to take

Paste into ChatGPT, Claude, Codex, or your tool of choice, and replace the bracketed bits. Run it when the work is big enough that the wrong model shape would waste real time.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. 🚀 SpaceX Just Pulled Off the Biggest IPO in History

SpaceX went public on Nasdaq under the ticker SPCX, capping a roadshow that opened on June 4 with pricing after the close on June 11 and a first trading day on June 12. The offering put up roughly 556.6 million shares at about $135 each, a raise near $75 billion that values the company around $1.75 trillion, which would make it the largest IPO ever. Unusually, Elon Musk pushed to reserve up to 30% of the shares for retail investors, far above the typical 5% to 10%. Starlink is the engine of the story: the satellite-internet unit now counts about 10.3 million subscribers and is the only consistently profitable part of the business, estimated at well over half of 2024 revenue. SpaceX also folded in Musk’s AI company xAI as a wholly owned subsidiary in an all-stock deal earlier this year. The Wall Street Journal framed the listing as a defining moment for Musk’s legacy. (Read More)

🫖 TEA For Thought: “This is a miracle. Like Musk said during the Nasdaq speech, you have to do something exciting, something that made you excited when you woke up.”

2. ⛓️ Meta’s Drafted AI Engineers Call Their Months-Old Unit a ‘Soul-Crushing Gulag’

Meta’s new Applied AI team, stood up only about three months ago and now roughly 6,500 engineers and product managers strong, has become a morale crisis. The core grievance is that Meta is involuntarily reassigning, or “drafting,” staff into the unit with little choice. In a leaked audio recording, Mark Zuckerberg defended the move by saying Meta employees are “significantly higher” in ability than outside contractors for the work. The team, led by 12-year veteran Maher Saba (previously head of the Reality Labs division that burned through $83 billion on the metaverse), spends its days generating coding puzzles and training datasets to sharpen Meta’s AI models, with up to 50 people reporting to a single manager. During a livestreamed employee-only presentation, someone hijacked the broadcast with an expletive-laden outburst. Workers told Wired “it’s literally the gulag” and that the work is “soul-crushing.” Separately, more than 1,600 employees signed a petition protesting systems that track clicks and keystrokes for AI training data. Zuckerberg admitted in a Friday memo that recent changes had “caused distress.” (Read More)

🫖 TEA For Thought: “This shows how hard it is to pivot even when you are a tech company yourself. No matter how big you are, and just because you are big, becoming AI-pilled is so much harder than you think. Big chances for small entrepreneurship.”

3. 🕵️ Two-Thirds of Office Workers Are Smuggling Their Own AI Into the Job

A Wakefield Research and PagerDuty survey of 1,250 office professionals found that about two-thirds use unauthorized AI tools at work, opening serious data-security holes. A striking 88% admitted feeding work information into public chatbots: 43% pasted in emails, 40% meeting notes, 34% customer data, and 31% financial information or confidential documents. Consequences exist but have not stopped it, as 53% were told to stop and 48% faced formal discipline. The motivations are telling: 39% would rather use an unsanctioned tool than have no AI at all, 33% wanted to avoid management scrutiny, and 30% were dodging restrictive policies. Even though 86% of organizations have an AI policy, 81% of employees believe leadership holds itself to a different standard, and 72% think they understand AI better than their own technical teams. PagerDuty CTO Tim Armandpour called the trend a “massive enterprise liability” and urged executives to “redirect that energy into proven platforms that offer governance and automation at scale.” (Read More)

🫖 TEA For Thought: “Well, whose fault is it when employees BYOA (bring your own AI) to work? If the enterprise rollout is so slow and so full of restrictions, how can they prepare employees to increase productivity and all that? If employees are willing to spend extra money out of their own pocket to buy an AI tool to do their job, then it is the employers who need to do some reflection: what are they failing to do that causes this?”

4. 💸 That $200 AI Subscription Is Quietly More Generous Than the Labs Let On

SemiAnalysis (@SemiAnalysis_) put the math behind AI subscriptions to the test, buying the $200-a-month plans from both Anthropic and OpenAI and running long coding tasks until they hit the weekly limits. The headline finding cuts against conventional wisdom: rather than the assumed roughly $2,000 of API tokens you might expect a $200 plan to cover monthly, the subscriptions are “actually far more generous” than people think. The flip side is the labs’ economics. Assuming a 75% gross margin on raw API usage, the analysis estimates that subscription margins end up significantly thinner than selling tokens through the API. Because openly degrading a subscription invites public backlash, SemiAnalysis predicts the labs will instead “withhold new features/models from subscription plans,” and floats the possibility that an upcoming model referred to as “Mythos” could arrive API-only. (Read More)

🫖 TEA For Thought: “Hopefully with competition the prices will come down. But oh man, how great it is right now to use subscriptions, think about the value of the tokens. The real question, though: how much value can you, as a developer, generate from these subscriptions?”

5. 🪙 The SEC Moves to Scrap a 20-Year-Old Rule Blocking Tokenized Stocks

The SEC proposed rescinding Rule 611 and Rule 610(e) from its national market system regulations, a change Galaxy’s head of research Alex Thorn called “one of the biggest unlocks yet for tokenized stocks.” Rule 611, the Order Protection Rule, bans “trade-throughs,” meaning an order on one venue cannot execute at a worse price than is available elsewhere, while Rule 610(e) limits how exchanges display competing quotes. Those rules quietly outlaw the way decentralized finance actually works: automated market makers (AMMs) fill orders against whatever the pool price is and cannot pause to route to a better quote, so Thorn notes a tokenized stock pool “would commit trade-throughs constantly and arguably be an illegal trading center.” The SEC, which launched “Project Crypto” in August 2025, plans to replace the rules with a “best execution” framework that could finally permit AMMs, and has opened a 60-day public comment period. The agency had previously postponed a tokenized-stock trading plan after exchanges raised execution concerns. (Read More)

🫖 TEA For Thought: “More liquidity for RWA (real-world assets) on the way!”

🛠️ Skill of the Day

The Subscription ROI Check: point it at one paid tool you are not sure about, and get an honest keep, downgrade, or cancel verdict based on the value it actually creates.

You are a blunt ROI analyst, not a salesperson and not my cheerleader. I want an honest verdict on whether a paid subscription is worth keeping.

THE TOOL: [NAME, E.G. "CHATGPT PLUS," "NOTION," "A $200/MO AI CODING PLAN"]
WHAT IT COSTS: [PRICE PER MONTH OR YEAR]
WHAT I ACTUALLY USE IT FOR: [THE 2 TO 4 THINGS YOU REALLY DO WITH IT]
HOW OFTEN I USE IT: [DAILY / WEEKLY / RARELY]

Do this in order:
1. Estimate the real value it creates for me each month (time saved at my hourly rate, money earned, or money avoided). Show your math and flag every guess.
2. Subtract the cost and give me the honest monthly net. Tell me if I am fooling myself about how much I actually use it.
3. Name one cheaper or free tool that would cover about 80% of my real use, if one exists.
4. End with a one-word verdict, KEEP, DOWNGRADE, or CANCEL, plus the single biggest reason.

Be skeptical. If my usage does not justify the price, say so plainly.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. Run it on every recurring charge you are not sure about.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

Elon Musk’s companies are not separate bets. They are one connected plan. XAI is the brain, Tesla is the body, X is the voice and data layer, Neuralink is the human interface, The Boring Company builds the underground infrastructure, and SpaceX is the transport that takes all of it beyond Earth. AI will eventually be free for everyone. When that happens, the real question is not what AI can do. It is what you will do with it.

Everyone is talking about the SpaceX IPO. But the IPO itself is not the story. The story is what it signals.

After watching Lewis Hong’s SpaceX VC interview, something clicked for me. SpaceX is not a rocket company anymore. It is infrastructure, the way the internet was never really just email. Rockets happen to be the vehicle, but what Musk is actually building is the ability to move mass, energy, and intelligence beyond Earth at a cost that was previously unthinkable. Reusability is the unlock. Once you can reuse a rocket the way you reuse a plane, the economics of everything in space changes.

Lewis made a point that I keep thinking about. Scientists were spending enormous effort building labs that simulate space conditions, trying to manufacture things that zero gravity makes easy. At some point, the obvious question becomes: why simulate it when you can just go?

That question is bigger than it sounds. It is the same logic behind everything Musk is building.

The Ecosystem: Brain, Body, Voice, Interface, Infrastructure

Here is what I did not fully appreciate until now. These companies are not separate bets. They are one system.

XAI is the brain, the intelligence layer. Tesla is the body, physical AI made real. Full Self-Driving is essentially a robot navigating the world, and the Optimus humanoid is just the next step in that direction. X, formerly Twitter, is the voice and the data layer. Hundreds of millions of people post on X every day, which makes it the world’s largest real-time feed of human thought. That data trains xAI. The brain needs a constant stream of information to learn from, and X provides it.

Neuralink is the interface. With human consent, a chip in the brain connects a person directly to this network of intelligence. It is not science fiction anymore. It is the bridge between human cognition and machine capability, a way for people to stay not just relevant but genuinely enhanced in a world where AI handles most of the cognitive and physical work.

And then there is The Boring Company. On the surface, it builds tunnels. But think further ahead. If humans are going to live on Mars or the Moon, the surface is not where we will be. Radiation, temperature swings, no atmosphere. We will live underground. The Boring Company’s entire capability points directly at that problem. Lewis Hong raised this connection, and I think it holds up.

All of it, XAI, Tesla, X, Neuralink, The Boring Company, SpaceX, points toward one destination: make humanity multi-planetary and bring science fiction to life.

The Space Economy: What Gets Built When the Cost Collapses

A new economy is forming around space, and most people have not noticed yet.

We are talking about data centers in orbit, powered by unlimited solar energy and cooled by the vacuum of space, connected to Earth through Starlink’s laser links. We are talking about zero-gravity manufacturing for materials that simply cannot be produced on Earth. We are talking about colonies on the Moon and Mars, not as distant fantasy but as the next logical step in a plan that is already being executed.

When Musk did not have enough chips, he built his own. Terafab is doing that now at a scale that dwarfs anything previously attempted. When he did not have enough compute, he built that too. Anthropic and Google have already signed deals with SpaceX to access that compute. And Musk is not overselling it. He is offering 90-day commitments, no more, because he knows the pace of change makes longer promises unreliable.

Eventually, robots will do the heavy work on Mars. Humans will direct, explore, and build. That is not a distant dream. It is the logical endpoint of what Tesla’s Optimus program is already working toward.

Eric Jorgensen, who wrote the definitive book on Elon Musk, noted something worth sitting with:

the world’s first trillionaire is not a banker, not a fund manager, not a politician. He is an engineer who builds things that did not exist before.

That matters. It is a signal about what the world is beginning to value, and what kind of work actually moves civilization forward.

When Musk spoke at the Nasdaq IPO, the excitement was not performed. He said it himself: what keeps you going is waking up every morning unable to wait to see what happens next. That kind of genuine drive is rare, and it shows in the results.

AI Is the Tool. The Universe Is the End.

Here is the part that people miss when they talk about AI.

Intelligence is becoming a commodity. The leading models are converging. Open source is catching up fast, and the gap between the best closed model and the best free model is shrinking every few months. At some point, and it may not be far off, capable AI will be available to anyone at essentially zero cost.

That does not weaken the Musk ecosystem. It actually makes it more valuable. When intelligence is free, the scarce thing becomes what intelligence enables: physical infrastructure, energy systems, transportation, manufacturing, space. You cannot download a rocket. You cannot fork a Mars colony on GitHub. The companies in this ecosystem are not selling intelligence. They are building the world that intelligence makes possible.

So the question shifts. Not what can AI do, but what will you do when AI can do almost anything?

AI is the tool, not the destination. The destination is something much older and much harder to define: understanding what humanity is, finding out what the universe is made of, and asking questions that no one has thought to ask yet. Musk has said it in different ways across many speeches. The universe is the answer. You just have to figure out the right questions.

When AI can answer almost everything, what will you ask?

That is the question worth thinking about.

Till next time, Cheers!

Subscribe now

Previous column articles can be found here:

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

Recommended watching:

Elon Musk’s Nasdaq IPO Speech

SpaceX Future Plan Presentation

Lewis Hong on SpaceX and the Space Economy

1. 🪙 Coinbase Hands Your Crypto Wallet to an AI Agent That Trades and Buys Its Own Research

Coinbase launched its own trading agents on Thursday, just days after Robinhood rolled out agents that trade for users. The agent connects to your main Coinbase account, or to a separate sandbox if you would rather not give it the keys, and can rebalance a portfolio, follow an investment thesis you hand it, or weigh in on a one-off trade. It works through Coinbase Advanced, the pro platform with TradingView charts, and can trade crypto spot markets and derivatives now, with equities and prediction markets planned next. The twist is payments: the agent runs on x402, the open payment protocol Coinbase launched last year with AWS, Anthropic, Circle, and Near, so it can pay for premium research APIs and on-demand compute with no login or subscription. It also runs inside ChatGPT or Claude through a Coinbase MCP server. Custom guardrails like maximum trade size and spending caps are coming, and the Financial Stability Board this week called for stronger safeguards on agentic AI in finance. (Read More)

🫖 TEA For Thought: “The day has come.”

2. 🔒 Microsoft Bars Its Own Employees From Anthropic’s Claude Fable 5 While It Ships the Same Model to Customers

Microsoft has restricted its own employees from using Anthropic’s newly launched Claude Fable 5 while its legal and compliance teams review how the model handles data. The sticking point is Anthropic’s retention policy for the Mythos class of models, which Fable 5 belongs to: prompts and outputs are kept for 30 days for safety monitoring, and anything flagged by Anthropic’s safety systems can sit on its servers for up to two years. That clashes with the enterprise data rules Microsoft has in place, and the company is worried staff could feed customer data, corporate details, or other confidential information into the model. Fable 5 is absent from the model picker for Microsoft’s internal GitHub Copilot, and workers have been told to lean on Microsoft’s own tools instead. The irony is that Microsoft rolled the very same model out to its GitHub Copilot and Foundry customers without hesitation. It also follows Microsoft moving its developers off internal Claude Code licenses toward GitHub Copilot. (Read More)

🫖 TEA For Thought: “It is hard to do the right thing, and sometimes when you push too far the result is not what you intended. So what is Anthropic actually trying to achieve here? Safety and security by controlling and censoring everything? And if so, who gets to audit Anthropic?”

3. 🧮 Ramp Starts Renting Out the AI Finance Engineers It Built for Itself

Ramp introduced Applied AI Solutions, a service that embeds Ramp engineers inside a customer’s finance team to build and ship AI workflows on the Ramp platform in a matter of weeks. The pitch grows straight out of Ramp’s own operation: its finance team leans heavily on production AI agents running on a “Finance Intelligence” layer that maps GL accounts, policies, and reporting logic to how the business actually works, letting Ramp run with a fraction of the headcount its scale would normally require. The new team runs a structured discovery process, connects data wherever it lives (ERPs, warehouses, cloud storage, even paper workflows), builds that same intelligence layer for the client, and deploys agents tied to the customer’s KPIs. Ramp says it is model-agnostic, routing each workflow to whichever model performs best on cost and accuracy across the 70,000-plus businesses and 200 billion dollars it processes annually. The backdrop: across Ramp’s customers, AI token spend is up 13x since January 2025, yet while 87% of CFOs call AI critical, only 21% report measurable results. (Read More)

🫖 TEA For Thought: “This is a smart business model. You build the team that implements the tech for yourself first, then you sell that exact capability to everyone else. Ramp is essentially packaging its own finance-specialized forward-deployed engineers and selling them as the product.”

4. 🍔 DoorDash’s New ‘Ask DoorDash’ Lets You Order Dinner From a Prompt or a Photo

DoorDash launched an AI chatbot called Ask DoorDash that lets people order food and groceries using plain-language prompts and photos instead of scrolling through restaurants and stores. You can tell it what you are in the mood for, drop a recipe link to pull the ingredients, or snap a photo of a cookbook page or grocery list, and it will assemble the cart with the right items and quantities, even prompting you to check whether you already have staples like sugar and butter. For restaurants, a query like “a filling dinner for a family of four” surfaces options with a short blurb on why each fits, and you can narrow it further to “kid-friendly vegetarian spots with mild options.” It also handles reservations, taking a request like “a table for two downtown for date night around 8 PM” and surfacing places with availability. The chatbot is rolling out on iOS in select regions and will reach more of the US in the coming weeks, joining similar moves from Uber Eats and Instacart. (Read More)

🫖 TEA For Thought: “A pretty good and authentic pivot. The only real question is whether the cost of image generation, LLM calls, and the rest actually pencils out.”

5. 🍎 Apple Shelves the New Siri for All of Europe With No Timeline to Bring It Back

Apple confirmed it will not ship its overhauled, AI-powered Siri as part of iOS 27 and iPadOS 27 in the European Union, blaming conflicts with the Digital Markets Act, and it gave no timeline for when, or whether, that changes. According to Apple, the DMA would force it to give any AI system “nearly unlimited access to a user’s device” plus the ability to act autonomously “without a user’s ongoing visibility and control.” Apple says it proposed a “Trusted System Agent” intermediary and an 18-month rollout plan, but the European Commission rejected its proposals. Craig Federighi said Apple hopes to bring Siri AI to the EU eventually but that the Commission’s “refusal to engage constructively” leaves it without a timeline. EU developers also cannot test the Siri AI features on the betas, and the new Siri and Apple Intelligence features will not ship in China either, citing regulatory work. EU users can still get the revamped Siri on macOS 27, visionOS 27, and watchOS 27. (Read More)

🫖 TEA For Thought: “The decline of Europe in a single headline: rare innovation, extreme regulation.”

🛠️ Skill of the Day

The Agent Guardrail: before you hand an AI agent your wallet, your inbox, or your production keys, make it design its own safety rails first, so you find the failure modes before they find you (ask Coinbase and the Financial Stability Board why this is suddenly the conversation).

You are a risk engineer, not my assistant. I am about to let an AI agent act on my behalf. Before it touches anything real, your job is to design the guardrails that keep a bad day from becoming a disaster.

THE AGENT'S JOB: [DESCRIBE WHAT THE AGENT WILL DO, E.G. "REBALANCE MY PORTFOLIO," "ANSWER CUSTOMER EMAILS," "DEPLOY CODE"]
WHAT IT CAN TOUCH: [ACCOUNTS, MONEY, DATA, SYSTEMS, PEOPLE]
WORST REALISTIC LOSS: [MONEY, DATA, REPUTATION, TIME]

Do this in order:
1. List the 5 ways this agent could cause real damage, ranked by how badly it would hurt, not how likely it is.
2. For each one, give me a concrete hard limit that would cap the damage (a spending ceiling, a read-only scope, a human approval step, a rate limit, a kill switch).
3. Tell me which single action this agent should NEVER take without a human pressing the button, and why.
4. Design the "panic test": one thing I can check each day in under a minute to know the agent has not gone off the rails.
5. Write the agent's operating rules as a short instruction block I can paste into its system prompt, written so a careful agent would refuse to cross the lines you set.

Assume the agent is competent but literal, and that anything you do not explicitly forbid, it will eventually try.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. Run it before you give any agent permissions, not after.

TEAHEE Moment

Stay sharp, stay informed. See you Sunday!

Subscribe now

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

1. ⚖️ Both Lawyers Used AI, So a Judge Canceled the Trial and Removed Everyone From the Case

In a federal contractual dispute in the Northern District of Mississippi, lawyers on both sides were caught citing nonexistent, AI-hallucinated cases, a situation one observer described as two clients effectively paying for an LLM to argue against itself. Senior U.S. District Judge Sharion Aycock issued a blistering sanctions order warning that “in an era of rampant unverified AI usage within the legal field, this case presents a prime example of the risk associated with serving as a rubber-stamp.” She paused the proceedings, canceled the trial, and disqualified all four lawyers involved. Two were barred from appearing before the court for two years, and all received fines between $1,000 and $3,500 depending on her assessment of their culpability for not verifying the AI’s output. Judges nationwide have grown increasingly frustrated with hallucinated citations, with a similar case ripping into New York lawyers just last week. (Read More)

🫖 TEA For Thought: “The verification, stability, and reliability are essential. But just because this happened does not mean AI is all bad. Everyone is racing to build enterprise AI that is safe, reliable, and durable.”

2. 💬 The EU Orders Meta to Open WhatsApp to Rival AI Chatbots, for Free

The European Commission ordered Meta to give rival AI chatbots free access to WhatsApp within five working days or risk a heavy fine, an interim measure in its ongoing antitrust probe. The investigation, opened in December, targets Meta’s policy of blocking AI providers other than its own Meta AI from the platform. The EU had warned in February that interim measures were coming, then rejected as unsatisfactory the access fee Meta introduced in April. Meta must now restore access on pre-October terms, with the order lasting until the investigation concludes or June 2029 at the latest. The Commission can impose a fine of up to 10% of Meta’s total annual turnover if the company intentionally or negligently defies the decision. Meta said it would appeal. (Read More)

🫖 TEA For Thought: “The automation that uses WhatsApp as an interface can start now!”

3. 🌊 China Switches On the World’s First Wind-Powered Underwater Data Center

China has begun operating what is being called the world’s first offshore wind-powered underwater data center, off the coast of Shanghai’s Lingang area. Located around 10 kilometers from shore and submerged roughly 10 meters below the surface, the facility is powered directly by a nearby offshore wind farm, which is estimated to supply 95% of the electricity for its 192 server racks arranged across four levels. It uses seawater for natural cooling through a circulating copper-pipe heat exchange system that reportedly cuts electricity consumption by 22.8%, needs no freshwater, and reduces land use by more than 90% compared with onshore centers. The site is currently running at 2.3 MW after completing its first phase in late May 2026, with a planned capacity of 24 MW, enough to power roughly 20,000 households. (Read More)

🫖 TEA For Thought: “Well, in CCP China, if they want to do anything, the environment is the last thing anyone has to worry about, as long as there is an outcome.”

4. 🤖 A Perplexity Study Says Agents Do 26 Minutes of Work Where Search Does 33 Seconds

A new paper from researchers using production data from Perplexity’s Search and Computer products measures what happens when AI shifts from answering questions to doing the work end to end. Using near-identical query pairs as natural experiments for the same underlying task, they found Computer performs 26 minutes of autonomous work per session versus 33 seconds for Search, automating the task decomposition and execution a Search user would otherwise orchestrate by hand. Autonomy also raised quality, with per-query dissatisfaction rates 55% lower on Computer, and on matched tasks it cut completion time from 269 to 36 minutes, lowering estimated time and cost by 87% and 94% respectively compared with humans using Search alone. The agent also changed what users attempted, with queries more often crossing occupational boundaries and bundling interdependent subtasks into a single request. (Read More)

🫖 TEA For Thought: “Isn’t this partially showing that AI does work better than humans, better quality with less money? When more and more ordinary people discover that AI is not just answering questions for them but can actually do it for them, the general rollout of agentic AI might be starting.”

5. 🚀 The Three Hard-Tech Moonshots Behind SpaceX’s $75 Billion IPO

SpaceX comes to market Friday with a reportedly oversubscribed $75 billion offering, but a sober look at the plan shows most of the upside riding on a risky bet: orbital data centers. Two independent analyses value the company well below its roughly $1.8 trillion banker assessment, with Morningstar at about $825 billion and NYU’s Aswath Damodaran at $1.2 trillion, characterizing the gap as essentially a call option on SpaceX’s ability to deliver space-based AI compute. The vision requires three near-impossible feats at once: a fully reusable Starship rocket, a brand-new US chip foundry (Terafab), and a satellite production sprint to roughly 556 AI satellites a month, about twice Starlink’s current rate, to hit an annualized gigawatt of space compute by the end of next year. Meanwhile SpaceX is already renting compute to ostensible rivals Anthropic and Google, raising the question of whether it wants to be a model-builder or a cloud provider. (Read More)

🫖 TEA For Thought: “Only gravity is law, the rest is just recommendation.”

🛠️ Skill of the Day

The Hallucination Hunt: before you trust anything an AI hands you, make it turn around and attack its own answer, so you catch the made-up parts before they cost you (just ask those four Mississippi lawyers).

You are a meticulous fact-checker reviewing a draft you did NOT write. Below is an AI-generated answer I am about to rely on. Your job is to find what is wrong or unverifiable, not to defend it.

THE ANSWER TO CHECK:
[PASTE THE AI OUTPUT HERE]

WHAT I WILL USE IT FOR: [E.G. A FILING, A CLIENT EMAIL, A DECISION, A POST]

Do this in order:
1. List every specific factual claim, name, number, quote, citation, or date in the answer.
2. For each one, label it: VERIFIED (you are confident and can say why), UNCERTAIN (plausible but you cannot confirm), or LIKELY FABRICATED (looks invented, e.g. a case, study, or stat that may not exist).
3. Flag anything that sounds authoritative but is the kind of detail AI commonly hallucinates: legal cases, citations, statistics, URLs, exact quotes.
4. Tell me the single claim that would do the most damage if it turned out to be false, and exactly how I can verify it in five minutes.
5. Rewrite the answer with every unverifiable claim either removed or clearly marked "needs verification."

Be blunt. If you cannot stand behind a claim, say so plainly.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. Great before anything goes out the door with your name on it.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. ⛓️ MetaMask Launches a Wallet Built for AI Agents to Trade DeFi on Their Own

Consensys launched MetaMask Agent Wallet on June 8, a self-custodial product that lets AI agents trade across decentralized finance without requiring human sign-off on every transaction. It supports token swaps, perpetual futures, prediction markets, and liquidity provisioning across Ethereum-compatible chains and Hyperliquid, and every transaction the agent initiates is simulated first and scanned for threats by Blockaid in real time. Risky transactions, or any that fall outside user-defined limits, trigger a two-factor approval, and users set their own daily spending caps and protocol allowlists before granting access. Two modes are offered: Guard Mode, which enforces all rules strictly, and Beast Mode, which reduces prompts while still requiring sign-off on anything flagged as malicious. MetaMask, which holds roughly 26% of the wallet market, says the product is framework-agnostic and already works with OpenAI Codex, Claude Code, OpenClaw, and Cursor. A full public release is planned for later in summer 2026. (Read More)

🫖 TEA For Thought: “MetaMask just joined the party!”

2. 💻 Cognition’s FrontierCode Asks Whether AI Can Write Good Code, Not Just Correct Code

Cognition introduced FrontierCode, a benchmark that measures whether AI models can write code a human maintainer would actually merge, not just code that passes tests. It was built with the maintainers of 36 flagship open-source repositories, who spent more than 40 hours per task defining what “mergeable” means in their codebase and grading along axes like correctness, test quality, scope discipline, style, and design. The team says it produces 81% fewer misclassification errors than benchmarks like SWE-Bench Pro, which only test functional correctness. The results show even the best models struggle: on the hardest 50-task “Diamond” subset, Claude Opus 4.8 scored just 13.4%, with GPT-5.5 at 6.3% and Gemini 3.1 Pro at 4.7%, though GPT-5.5 used up to 4x fewer tokens. Cognition is keeping the tasks private to avoid contamination but is opening the evaluation to model creators. (Read More)

🫖 TEA For Thought: “I am struggling to understand this one. If writing good code is about efficiency or scalability, I get it. But writing good code purely for the sake of good code, what is the point? Code is a tool after all, and the purpose of a tool is to solve problems.”

3. 📵 The FCC Moves to Effectively Kill Burner Phones

The Federal Communications Commission wants to legally require U.S. telecoms to collect a government-issued identification number and physical address from essentially all new and renewing phone customers, a change 404 Media reports would make it nearly impossible to buy a so-called burner phone not tied to your identity. The FCC frames the data collection partly as a way to fight scammers, and would also require carriers to gather extra details on business and foreign customers, such as the intended use case of bulk purchases and IP addresses. Privacy advocates warn the rule would hit low-income people, domestic abuse survivors, journalists, and anyone privacy-conscious, with carriers storing sensitive personal data that creates new cybersecurity risks. The ACLU’s Jay Stanley compared it to measures in authoritarian countries where registration is required to get a mobile phone, saying “we never thought that would happen here.” (Read More)

🫖 TEA For Thought: “This is pretty concerning. It reminds me of how blockchain first came out, as a way for people to transact without handing their identity over to a central authority. Now the trend seems to be heading in the opposite direction.”

4. 🍎 Apple Plays Catch-Up on AI at WWDC 2026

Apple’s WWDC 2026 keynote, the last under CEO Tim Cook before John Ternus takes over September 1, leaned on fixes before features as the company tries to reassert itself in a competitive AI landscape. The headline was an overhauled Siri, now more conversational and powered by Google’s Gemini models under the hood, housed in its own standalone app, with Apple stressing that “privacy in AI is non-negotiable.” Apple Intelligence gained cross-app context awareness, AI reply suggestions in Messages, and a Phone app that can pull context mid-call, while the Photos app added AI editing tools like spatial “Reframe” and “Extend.” Other updates included opt-in rollbacks to the controversial Liquid Glass design, natural-language Shortcuts creation, systemwide dictation, new parental controls, App Store subscription bundles, and perimenopause support in Health. iOS 27 will reach devices as far back as the iPhone 11, and developer files hinted at a possible foldable iPhone down the line. (Read More)

🫖 TEA For Thought: “I was not going to include this, since it is the headline everyone is already talking about. But the keynote video is brutal to watch. Is that why Apple turned off the comments, afraid of the reaction? In the era of AI, staying stagnant means dying, and it is just brutal to watch.”

5. 🛰️ The Pentagon Adds Alibaba, Baidu, and BYD to Its Chinese Military List

The Pentagon expanded its “1260H list” of companies it considers tied to China’s military or defense industrial base, adding major tech and industrial names including Alibaba, Baidu, carmaker BYD, drug developer WuXi AppTec, robotics firms RoboSense and Unitree, and memory-chip makers CXMT and YMTC. The roster has now grown to 188 entities, up from 134 in the prior official revision, spanning AI, electric vehicles, robotics, and biotechnology. The designations do not impose sanctions directly, but the Defense Department will be barred from contracting with listed companies later this month, and from buying their products through third parties starting in June 2027. The move comes just weeks after President Trump met Xi Jinping in Beijing and the two agreed to a trade truce, underscoring the persistent tension between commercial thaw and Washington’s security concerns over Chinese technology. (The WSJ reported the expansion; corroborated by CNBC, NPR, and The Washington Post.) (Read More)

🫖 TEA For Thought: “No business can be totally independent from CCP control. Like Manus, you think you are, until you are not, and you have to raise a billion dollars to pay the acquisition fee.”

🛠️ Skill of the Day

The Delegation Audit: figure out what to hand off to an AI agent and what to keep for yourself, before you automate the wrong thing.

You are a pragmatic operations advisor. Help me decide what to delegate to an AI agent versus keep doing myself.

WHAT I DO: [LIST THE TASKS, PROJECTS, OR RESPONSIBILITIES ON YOUR PLATE]
MY GOAL: [WHAT YOU WANT MORE TIME OR ENERGY FOR]

Do this in order:
1. Sort my tasks into four buckets: automate now (repetitive, rule-based, low-risk), delegate with review (judgment needed but checkable), keep myself (high-stakes, relational, or creative), and drop (not worth doing at all).
2. For each "automate now" item, name the single biggest failure mode if an agent got it wrong, and the cheapest guardrail to catch it.
3. Point out one task I am probably holding onto out of habit that I could safely hand off.
4. Give me a one-week experiment: the single best task to delegate first, and how I will know if it worked.

Be concrete and honest. If I have not given you enough detail on a task to judge it, ask.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. Great for anyone staring at a long to-do list wondering where an agent actually fits.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. ⛓️ The 10 Crypto Sectors Most Likely to Survive and Grow Through 2026

Pink Brains breaks down 10 crypto sectors it believes can keep growing through the second half of 2026, arguing that most altcoins will never reclaim their 2021 highs and that measurable usage, not narrative, now drives capital. Topping the list are two AI-crypto categories: agentic finance, where onchain trading agents gained real transactional authority through new wallet standards like EIP-7702 and Base’s AgentKit, and the “picks and shovels” compute layer (Bittensor, Render, Akash, Grass), where the broader AI-crypto sector roughly tripled to $22 to 27 billion by May 2026. The rest of the list runs through prediction markets (a record $28.4 billion in monthly volume in May), perp DEXs, onchain vaults, RWA-backed DeFi, privacy coins, crypto cards, stablecoin rails, and cultural platforms like Pump.fun. The throughline: the old playbook was buying a narrative early, the new one is spotting real usage early. (Read More)

🫖 TEA For Thought: “The intersection between AI and blockchain has never been closer.”

2. ⚖️ New York Starts Policing AI Performers in Ads on June 9

New York’s synthetic performer law (S.8420-A / A.8887-B, amending General Business Law 396-b) takes effect June 9, 2026, and requires advertisers to conspicuously disclose when an ad uses an AI-generated or computer-created digital performer made to look like a human. It covers synthetic performers in primary or background roles, but exempts recognizable replicas of real, identifiable people, expressive works like films and games, audio-only ads, and AI used purely for translation. Ad creators who knowingly include a synthetic performer must label it, and publishers must fix or remove undisclosed ones within 5 business days of notice. Civil penalties run $1,000 for a first violation and $5,000 for each one after, enforced by the state with no private right of action. (Read More)

🫖 TEA For Thought: “Alert to any business that resides in New York!”

3. 🤖 Report: xAI Trained Its Coding Models on Claude for Months

The Information reports that xAI spent months training its coding models directly on the outputs of Anthropic’s Claude. After Anthropic revoked xAI’s official access in January 2026, engineers allegedly kept tapping Claude through personal accounts and an intermediary service called Blackbox AI. The reporting also describes turmoil inside xAI: the pretraining team shrank to fewer than five people, four Grok code leads and several co-founders departed within months, and an employee accidentally deleted critical training data, costing two to three weeks of work. Musk has previously acknowledged in court that xAI “partially” used OpenAI models for early Grok training, calling it standard industry practice. He is now renting out GPU capacity through SpaceX to Anthropic and Google as a stopgap. (Read More)

🫖 TEA For Thought: “The terafab is one of Musk’s true leverage points. No matter who you are, you have to kneel and make a deal with those who can provide the compute.”

4. ⚛️ America’s First Advanced Reactor Goes Critical in Idaho

Antares Nuclear has become the first company to bring an advanced reactor to criticality under the U.S. Department of Energy’s Reactor Pilot Program. Its Mark-0 demonstrator passed a zero-power criticality test at Idaho National Laboratory on June 4, 2026, meaning its nuclear reactions became self-sustaining (it does not yet generate power). The reactor runs on TRISO fuel, in which uranium oxide particles are wrapped in layers of carbon and ceramic for high-temperature safety. The milestone answers Executive Order 14301, which directed DOE to get at least three advanced reactors critical by July 4, 2026. The compact design is being eyed for military and remote-site microreactor use, an early step toward small, modular, lower-carbon nuclear power. (Read More)

🫖 TEA For Thought: “SMR might be the future of green power. Small, modular, clean.”

5. 🦾 Why Robotics Is Still Waiting for Its Big Theory

In a widely shared essay, roboticists Stefanie Tellex and David Watkins argue that robotics is still a “pre-paradigm” field, borrowing Thomas Kuhn’s term for a science that has not yet agreed on an organizing theory. They note the field juggles competing frameworks (subsumption architecture, SLAM, POMDPs, neural networks, diffusion policies) with no grand unified theory of embodied intelligence. The current industry bet, placed by companies like Physical Intelligence and Tesla, pairs large datasets with vision-language-action models and reinforcement learning, yet the authors say even deep learning lacks a full theory of why it works. They compare the moment to chemistry before Lavoisier, when Priestley isolated oxygen but still explained it through phlogiston, and argue robotics is waiting for its own paradigm shift. (Read More)

🫖 TEA For Thought: “Rushing to find the next game to play.”

🛠️ Skill of the Day

The Decision Stress-Test: pressure-check any big decision before you commit, so you find the holes while they are still cheap to fix.

You are a sharp, friendly red-team advisor. I am about to make a decision and I want you to stress-test it, not cheerlead it.

MY DECISION: [DESCRIBE THE DECISION AND WHY YOU ARE LEANING THIS WAY]
CONTEXT: [BUDGET, TIMELINE, WHO IS AFFECTED, WHAT IS AT STAKE]

Do this in order:
1. Restate my decision in one sentence so I know you understand it.
2. Give the 3 strongest reasons it could be right.
3. Give the 3 strongest reasons it could go wrong, including the failure mode I am most likely ignoring.
4. Name the single assumption that, if false, breaks the whole plan, and tell me the cheapest way to test it this week.
5. Offer one smaller or reversible version I could try first.

Be concrete and honest. If my reasoning is thin, say so plainly and tell me what is missing.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. Great before a hire, a big purchase, a launch, or any bet you cannot easily undo.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. 🤖 Amazon’s Proteus Robot Goes Fully Autonomous and Takes Orders in Plain English

Amazon has upgraded Proteus, its heavy-duty warehouse robot, into a fully autonomous machine that workers direct with plain conversational language instead of custom software. Amazon Robotics VP Scott Dresser put it simply: you tell it what needs to be done, and it figures out the priority, the route, and the timing. Proteus hauls heavy carts, moves containers as they arrive, and shuttles them between workstations, now roaming the whole warehouse rather than just the loading dock. It joins sibling robots Vulcan, which is touch-sensitive, and Stark, which handles smaller containers with precision. Amazon has deployed over 1 million robots so far. The AI-powered Proteus is still in lab testing, with a European rollout planned for the first half of 2027 alongside 25,000 new warehouse hires there. (Read More)

🫖 TEA For Thought: “A reminder that robots do not have to look exactly like human beings. As long as it can do what it does in a physical form, it is a robot. Amazon is clearly leading on warehouse robots, because this one does exactly what it is asked to in human language.”

2. 🧪 ViBench Launches as the First Benchmark for End-to-End Vibe Coding

ViBench is the first open-source benchmark that scores AI agents on end-to-end vibe coding, turning a plain-language request into a working web app rather than testing isolated snippets of code. Its tasks come from real production traces across 15 applications and are defined purely by user-facing requirements, with no implementation constraints or reference solutions, so the only thing that counts is whether the finished app does what was asked. An adaptive evaluator drives each generated app through a human-authored test plan using browser automation, reaching high agreement with human experts and exposing where today’s agents quietly break down. It was built by Hung Tran, Langston Nashold, Rayan Krishnan, Antoine Bigeard, and Alex Gu, and published at the ACM Conference on AI and Agentic Systems (ACM CAIS) in 2026. (Read More)

🫖 TEA For Thought: “This is cool. As models get smarter and tokens get cheaper, humans do not need to understand the fundamentals of coding, as long as whatever is being built actually works to solve the problem.”

3. 💸 The Token Bill Comes Due as Companies Scramble to Rein In AI Costs

As AI coding agents spread, companies are getting hit with runaway token bills even as the price per token keeps falling. TechCrunch reports that Uber burned through its entire 2026 AI coding budget by April, Priceline saw a Cursor renewal jump 4 to 5 times, and one company ran up a $500 million Claude bill after forgetting to set usage limits. Per-developer token use rose 18.6 times in nine months, and Goldman Sachs projects global token usage will multiply 24 times by 2030. Faros AI CEO Vitaly Gordon relayed one CTO’s dilemma about an engineer who spent $40,000 on tokens in a single month. A wave of startups (Pay-i, Paid, Factory) and incumbents (Ramp, Datadog) now sell token-cost controls, and the Linux Foundation is launching a Tokenomics Foundation in July 2026. (Read More)

🫖 TEA For Thought: “’One of my engineers spent $40,000 on tokens last month, and I genuinely do not know whether I should stop him or go tell everyone else to be like him.’ Oooh, it is tough. But if more tokens can make you ship 10x faster, would you do it? There is always a tradeoff between time and money, always, both personally and in business.”

4. 🎭 Founders Fund Launches a Mafia Game Show Starring Sam Altman and Palmer Luckey

Founders Fund, the venture firm co-founded by Peter Thiel, has launched a card-game show called MAFIA the GAME, based on the party game Mafia, in which tech luminaries play to deceive and survive. The debut episode features OpenAI CEO Sam Altman, Anduril founder Palmer Luckey, biohacker Bryan Johnson, and Signal founder Moxie Marlinspike. It is hosted by Mike Solana, the Pirate Wires editor and Founders Fund chief marketing officer, who said he is bored with typical VC content and wanted a more interesting way to get to know someone. The show is pitched as an ongoing series with more episodes planned, and TechCrunch covered its launch on June 4, 2026. (Read More)

🫖 TEA For Thought: “Very fun to watch!”

5. 🏛️ S&P 500 Refuses to Fast-Track SpaceX, OpenAI, and Anthropic

S&P Global has declined to change its rules to let SpaceX, OpenAI, and Anthropic join the S&P 500 quickly after they go public, ruling that exceptions to its financial-viability, seasoning, and Investible Weight Factor requirements should not be granted based on market value alone. All three must first post four consecutive quarters of positive GAAP earnings. SpaceX reported a $4.94 billion net loss in 2025 on $18.67 billion in revenue, and neither OpenAI nor Anthropic is profitable, which pushes likely inclusion to late 2027 at the earliest. Bloomberg Intelligence estimates the three companies will collectively miss out on roughly $27 billion in forced passive-fund buying that fast index entry would have triggered. The decision was reported in early June 2026. (Read More)

🫖 TEA For Thought: “To me, it sounds like someone’s cheese is being touched, and it is all about control and politics. On one hand, the pre-IPO stocks of these three companies are trading like crazy and everyone wants in. On the other hand, the S&P 500 is saying no, it is too risky, we will not put money in for your own good. AI is disruptive in lots of ways, and the way traditional finance runs and operates might be one of the things getting disrupted.”

🛠️ Skill of the Day

The 20-Minute Expert: turn any article or topic into a crash course that actually sticks.

You are a patient expert tutor. I want to truly understand the material below, not just skim it. Teach it to me so it sticks.

MATERIAL OR TOPIC: [PASTE AN ARTICLE OR DOC, OR JUST NAME A TOPIC]
MY LEVEL: [complete beginner / some background / rusty]

Go in this order:
1. The big picture in 5 plain sentences: what this is and why it matters. Define every jargon word inline the first time you use it.
2. The 5 core ideas I must hold in my head, each in one line.
3. One concrete example or analogy for each core idea.
4. The 3 mistakes or misunderstandings beginners usually have here.
5. Five quick questions that test whether I really get it, hardest last. Wait for my answers before you reveal the solutions.

Keep it concrete and friendly. If the material is thin or you need more context, tell me exactly what to paste or read next.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. Great for getting up to speed fast before a meeting, an interview, or a new project.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

I was the thing deciding, every single step, whether to continue. The moment I walked away, it stopped.

Then I watched Boris Cherny, the head of Claude Code at Anthropic, say he does not really do that anymore.

In a recent talk, he said Claude Code went from writing 10 to 20 percent of his code to replacing his editor entirely. He uninstalled his IDE after not opening it for a month. And the part that stopped me: instead of prompting by hand, he now writes loops. Automated workflows that prompt Claude and decide what to build next. His job shifted from coding to orchestrating. He called it “the golden age of the generalist,” where designers, chiefs of staff, and finance people all ship real software.

The talk is here, it is worth 30 minutes.

The strange part: I had already built the pieces

When I looked back at my skills folder, I found something funny. Over the last two and a half months, I had quietly built a whole set of small skills, each one solving a single problem I kept hitting. I just never saw them as one thing.

The missing piece was about trust

There was one problem none of those tools solved: how do you know an AI’s output is actually good?

Regular software is a vending machine. Same input, same output, “all tests pass” means done. But anything with a language model inside it is different. The same prompt gives different answers across users, sessions, and model updates. What you shipped is not a fixed thing. It is a range of behaviors.

A few weeks ago I read an article by Jeff Gothelf called “What ‘done’ means when you’re shipping AI features.” One line stuck: if “done” is the language of a build culture, “calibrated” is the language of a learning culture. An AI feature is never simply done. It is calibrated. You have decided the variance you will accept, planned for the ways it fails, and you keep watching after launch.

Two days later I built a skill for exactly that (ai-done). It is the piece that decides whether the AI part of a product is good enough, not by a checkbox, but by measuring the spread of its output against a bar I set.

Then I tied them all together. I call it Ship It

Ship It is not a new capability. It is a thread.

You name a feature once, and it runs all of those skills in order, on its own. It asks me up front for anything only I can give, like a password or a login. It writes the checklist first. It builds in a loop while fixing its own bugs. It runs the health check. It calibrates anything with AI in it. Then it opens the real thing in a browser and QAs it, like a user would, until it actually works. It only comes back to me when it is ready to test.

The mechanism underneath is almost silly in how simple it is. There is a small script that fires every time Claude tries to end its turn. If the checklist is not all green yet, the script hands Claude the same instruction again and says “keep going.” Claude looks at the files it already changed, sees where it got to, and does the next piece. Try to stop, get handed the task again. That is the loop that prompts itself.

The real lesson: the loop is the easy part. Anyone can write “keep going until done.” The hard, valuable part is the verifier, the checklist that defines “done” so precisely a machine can check it. Without it, a self-prompting loop either runs forever or lies that it finished. With it, you can walk away. The verifier is the product, not the loop.

Where this is going: the cloud, and the context problem

Boris does not just run loops. He runs them in the cloud, on a schedule, without his laptop staying on. That is where I want to go, and there is an honest catch worth naming.

The limit on a long autonomous run is the context window, the model’s short-term memory. Let one run go too long and it fills up, starts summarizing itself to make room, and slowly gets dumber. There is no magic “context is full, hand off cleanly” button yet. That part is genuinely unsolved.

But here is the trick that makes it not matter. You do not run one giant session. You run many short ones. Anthropic has a feature called Routines: a scheduler that runs Claude Code on their own machines, pulls your project from GitHub, and fires on a timer, your laptop closed. So you set it to run, say, every hour. Each run is a fresh, empty mind. It opens the checklist, sees what is still unchecked, does a bounded chunk of work, commits it, and stops, well before its memory fills. The next hour, a brand-new run reads the same checklist and continues.

It is two loops stacked. The inner loop, inside one run, keeps prompting itself until the checklist passes or it hits a turn limit you set. The outer loop, the hourly schedule, keeps starting fresh runs until every item is finally ticked. Because no single run is ever long enough to fill the context window, the wall is never hit. The checklist sitting in the repo is the thread that ties all those short runs into one long job.

The reason it works is the same reason Ship It works at all: the memory lives in files, not in the conversation. A fresh session is never starting from zero. It just reads the list and keeps going.

The shift was never a tool

What changed for me was not a clever skill. It was a stance.

Stop being the loop. Start designing the verifier, the thing that knows when the work is truly done, and let the loop run against it.

The golden age of the generalist is real. And the wild part is, you might already have built most of the pieces. You just have to thread them.

Till next time, Cheers!

Subscribe now

Previous column articles can be found here:

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

1. 🫀 MIT’s Ultrasound Sticker Steadies the Heart Without Surgery

MIT engineers have built a noninvasive pacemaker that steadies the heart with ultrasound instead of surgery. The prototype is a sticker about the size of a postage stamp, wired to a pocket-sized box of batteries and electronics. It runs on sonogenetics: a one-time genetic tweak makes heart cells respond to ultrasound, and tiny transducers then send waves that trigger the ion channels behind each contraction. Led by professor Xuanhe Zhao with first author Chen Gong, plus collaborators at USC, Harvard, and UCLA, the team tested it on engineered human cardiac cells and in rats. Roughly 3 million U.S. adults live with implanted pacemakers today, all of which require surgery and direct contact with the heart. The study ran in Nature Biomedical Engineering on June 2, 2026, and the next goal is a single sticker that both images and stimulates the heart. (Read More)

🫖 TEA For Thought: “Portable hardware for health is the future, not just for general wellness but for specific diseases. After all, we need something to help us collect the data to upload a human!”

2. 💸 Teradata Tells Staff: No Raises This Year, the Budget Is Going to AI

Teradata, a cloud-analytics firm with about 5,100 employees, told staff in a January 2026 internal memo not to expect an annual salary raise this year, redirecting the compensation budget toward AI investments, according to Business Insider. The company casts autonomous AI and knowledge tools as the foundation of its next era. The memo drew attention because regular merit raises were frozen while the executive team still received increases. Steve McMillan is the president and CEO steering the shift. The decision became a flashpoint in a broader 2026 debate over how companies are reprioritizing payroll toward AI spending. (Read More)

🫖 TEA For Thought: “The money has to come from somewhere, so it is understandable. But with inflation, if your salary stays the same every year, you are actually earning less, because your purchasing power keeps shrinking.”

3. 🪄 Cash App Turns Tap-to-Pay Into a $25 Magic Wand

Cash App has launched a $25 tap-and-pay device shaped like a magic wand, inspired by a social trend of people hiding contactless cards inside homemade wands. The wand links to the Cash App Card and works anywhere Visa tap-to-pay is accepted, with no minimum balance, instant spend alerts, and the option to lock or deactivate it from the app if lost. Built by Block (Cash App’s parent), it went on sale Thursday, June 4, 2026, with more tag types, limited-edition drops, and permanent versions due by summer. Block hardware lead Thomas Templeton said the aim is to make payments visible and social, unlike invisible digital wallets and buried physical cards. The push targets Gen Z, following Cash App’s teen accounts and kids’ debit cards. (Read More)

🫖 TEA For Thought: “Imagine a marketplace for customized wands. This is cute!”

4. 🪙 Tether and Fasset Launch a Gold-Backed Rewards Card

Tether has teamed up with digital bank Fasset on a Visa rewards card tied to XAUT, its gold-backed token. Cardholders pay in fiat at any Visa merchant and earn up to 6% cashback in XAUT, with an automatic round-up that invests spare change into tokenized gold and rewards deposited in real time to a Fasset wallet. XAUT and Tether’s USDT can be converted to fiat at the point of sale. XAUT carries a market cap near $2.6 billion and is backed by about 24 tons of physical gold, inside a tokenized-gold market now worth over $5.3 billion. Tether has committed up to $1 million in XAUT rewards. CEO Paolo Ardoino said gold has long been a store of value, not a medium of exchange, and that this changes that. Fasset, which operates across Asia and Africa, is targeting emerging markets. (Read More)

🫖 TEA For Thought: “Tether is really doing something here.”

5. 🔁 Claude Code’s Boss Now Writes Loops, Not Code

In a WorkOS-hosted Acquired Unplugged talk, Anthropic’s head of Claude Code, Boris Cherny, laid out how AI is reshaping software work. He said Claude Code went from writing 10 to 20% of his code to replacing his editor entirely, and that he uninstalled his IDE after not opening it for a month. Instead of prompting by hand, he now writes loops, automated workflows that prompt Claude and decide what to build next, shifting his job from coding to orchestrating agents. He shared internal Anthropic figures: per-engineer productivity up 200%, over 90% of the Claude Code team’s code now written with Claude Code, and new-hire ramp-up down from weeks to about two days. He called it the golden age of the generalist, where designers, chiefs of staff, and finance folks all ship code. Asked what stays uniquely human, his answer was values: teaching the model to be good, the way we teach kids. (Read More)

🫖 TEA For Thought: “I am trying really hard to figure out how you create a loop where Claude Code keeps prompting itself, not just repeating the same prompt, but adapting based on errors and goals. Any insights?”

🛠️ Skill of the Day

The Self-Correcting Loop: make AI grade its own work against your goal and fix it before you ever see it.

You are my self-correcting assistant. I will give you a task and what a great result looks like. Do not show me your first attempt.

TASK: [describe what you want done]
DEFINITION OF DONE: [list 3 to 5 concrete things a great result must have]

Before you reply, work this loop silently:
1. Write a first draft of the task.
2. Score the draft from 1 to 10 on each item in the Definition of Done, and name the single weakest point.
3. Rewrite to fix that weakest point.
4. Repeat steps 2 and 3 until every item scores 8 or higher, or you have done 3 rounds.

Then show me only three things: the final result, the final scores, and one line on what changed between your first and last version. If you cannot reach an 8 on something, say so and tell me exactly what you need from me.

Paste into ChatGPT, Claude, or your tool of choice, and replace the bracketed bits. It is the hand-run version of the self-prompting loops in today’s last story.

TEAHEE Moment

Stay sharp, stay informed. See you Monday.

Subscribe now

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

1. 🔐 The Hackers Are Agents Now, And They Want Your AI Bill

Vercel detailed a new class of attack it calls inference theft. Attackers wrap a company’s AI endpoint in an OpenAI- or Anthropic-compatible adapter, then route traffic through residential proxy networks so the stolen model calls can be resold and used inside ordinary AI clients. The economics are lopsided: a frontier model call can cost around 2 dollars, while the HTTP request to trigger it costs a fraction of a cent. In one example, a forked coding agent turned Chipotle’s support chatbot into a free OpenAI-compatible endpoint, then openly recruited others to do the same to Home Depot, Lowe’s, Target, and Starbucks. On April 12, 2026, Vercel’s own docs assistant was hit with traffic spiking to 1,300 requests per minute and a projected run rate above 10,000 dollars a day. Vercel stopped it with BotID deep analysis (powered by Kasada), checking every single request instead of once per session, and blocked over 10,000 bot requests within minutes. (Read More)

🫖 TEA For Thought: “Now the hackers are also agents. Agent identity is something every website needs to figure out. The future of the economy is machine to machine. Everyone and everything is going to be agent-facing. Period.”

2. 🔌 AI Isn’t Plateauing, It’s Disappearing Into The Walls

Tech writer Om Malik pushes back on the popular line that AI is having its “iPhone moment,” arguing the analogy is wrong because it implies a plateau. His counterpoint is blunt: capability is still accelerating, but the conversation is shifting from AI as a hot topic to AI as invisible infrastructure. His preferred analogy is DWDM, the optical multiplexing technology that quietly multiplied internet capacity through fiber while vanishing from public conversation. Nobody mentions DWDM on a video call, yet it makes the call possible. Malik sees AI heading the same way, getting embedded inside cameras, chips, hospital monitors, and contracts until people stop noticing it is there. He argues most of the value will not accrue to the foundation model labs but to companies that build indispensable applications on top, much as Google and Netflix exploited cheap optical bandwidth. His 2028 prediction: the question stops being “which AI” and becomes “what does this do that it did not do before.” (Read More)

🫖 TEA For Thought: “AI isn’t getting boring. It’s just becoming a utility, like electricity or the internet backbone, that powers our world from behind the scenes.”

3. 🚀 YC’s Spring 2026 Batch Is Small Teams Shipping Coworkers

New Economies broke down Y Combinator’s Spring 2026 batch of more than 190 startups, and the headline is how lean the teams are. Nearly one in five companies has a single founder, and the median core founding team is just three people. The batch skews heavily B2B at 62 percent, with more than 110 B2B startups, followed by industrials, fintech, and healthcare. Around 60 percent of the one-line pitches mention AI or agents, and the framing has moved from shipping features to “shipping coworkers, and the tools those coworkers need.” Some categories shifted sharply: legal tech collapsed from seven startups in the previous batch to just two, hinting at saturation, while defense tech surged with strike drones, submarine drones, and aerial sensing systems. Healthcare stayed flat but tilted toward AI for clinicians and operations rather than biotech, and robotics plus industrials together make up roughly 12 percent of the batch. (Read More)

🫖 TEA For Thought: “Nearly one in five companies has a single founder, and the median core founding team is just three people. The leverage one person can now hold is wild.”

4. 🛒 Amazon Will Show You AI Products That Don’t Exist

Amazon said on Wednesday, June 3, 2026, that it will start showing AI-generated product images inside its shopping app, surfaced right below search autocomplete suggestions. The idea is to help shoppers who have something in mind but lack the exact words: type “blue gingham dress” and Amazon will render variations with different sleeve lengths and styles to steer you toward a better search. The company frames it as a visual search aid layered on top of its catalog of real listings, alongside earlier AI additions like review summaries, audio product descriptions, and the Alexa shopping chatbot. TechCrunch’s Sarah Perez is skeptical, calling it “somewhat bananas for a retailer to make up fake products as a way of guiding users,” and warning that shoppers may believe they can actually buy the invented item when Amazon already has real photographs it could show instead. (Read More)

🫖 TEA For Thought: “You can clearly tell Amazon is busy trying, exploring, and searching for product-market fit. No harm in trying. Ship it, and if it doesn’t work, figure it out and keep going. Every company is an AI startup now, including the big techs.”

5. 🔀 Perplexity Computer Will Split Your Work Between Laptop And Cloud

Perplexity is adding what it calls hybrid agentic inference to Perplexity Computer, its agentic system for handling real-world tasks on your behalf. The feature automatically splits each task between a compact model running locally on your device and more powerful frontier models in the cloud. A small on-device model decides, task by task, when sensitive information such as financial records, health data, and personal files should stay local, while heavier work that needs advanced capability runs on Perplexity’s servers. The pitch is that you never pick local or cloud up front, since the system routes the work on its own as it goes. It is slated to arrive in July 2026. The move fits Perplexity’s broader strategy of competing through orchestration rather than owning a foundation model, betting that the smartest harness wrapped around many models can beat any single one. (Read More)

🫖 TEA For Thought: “Perplexity is coming on really strong. No model of its own, but all harness, and it is really good at leveraging that for everything agentic.”

🛠️ Skill of the Day

The Analogy Engine: explain or pitch anything with an analogy that instantly clicks.

You are a master communicator who explains complex things with sharp,
memorable analogies. I will give you a product, idea, or concept that is
hard to explain. Help me make it instantly understandable.

Here is what I need to explain:
[DESCRIBE YOUR PRODUCT, IDEA, OR CONCEPT, AND WHO THE AUDIENCE IS]

Do the following:
1. State the core function in one plain sentence, no jargon.
2. Give me 5 analogies from different angles: one everyday, one
   historical, one from another industry, one "the X of Y" positioning
   line, and one slightly unexpected.
3. For each, name what it gets right and where it breaks down, so I do
   not oversell it.
4. Pick the single strongest analogy for my audience and explain why.
5. Write one tight sentence I could drop into a pitch, a landing page, or
   an intro using that analogy.

Keep it concrete. Avoid tired clichés unless one genuinely fits best.

Paste into ChatGPT, Claude, or your tool of choice, drop in anything hard to explain, and you walk away with a line that finally lands.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. 🚢 Shipping AI Means You Are Never Really “Done”

Product coach Jeff Gothelf argues the old definition of “done” falls apart for AI features. Traditional software behaves like a vending machine: the same input returns the same output every time, so “done” was the moment the tests passed and the spec was met. AI features instead ship a distribution of behaviors, where one prompt can yield different results across users, sessions, and model updates, so “all tests pass” still matters but no longer means done. Gothelf offers three fixes: write acceptance criteria as distributions (for 80% of inputs the response meets a quality bar, the other 20% fail gracefully), build the failure-triage playbook before launch instead of after complaints pile up, and set a tripwire metric (error rate, hallucination rate, off-tone complaints) with a rollback you have actually rehearsed. His reframe: done is no longer a checkbox, it is a stance on the variance and failure modes you choose to accept. (Read More)

🫖 TEA For Thought: “In the AI era, you can’t just check a box and walk away when you launch a feature. You are never truly ‘done.’”

2. 📊 The Top AI Adopters Aren’t All Big Tech (Energy Made The List)

A new open-source index from the AI-Driven Enterprise Institute (AIDE) ranks how far S&P 500 companies have actually adopted AI, scoring each up to 100 across four areas: literacy, advocacy, orientation, and implementation. It draws only on public data such as earnings call transcripts, job postings, and patent filings, so the ranking is not self-reported. Chipmaker Nvidia earned a perfect 100, joined at the top by Amazon, Meta, and, more surprisingly, energy producer SLB (formerly Schlumberger). Walmart (95.84) came next, followed by utilities AES (95.46) and NextEra Energy (95.44), with Chevron, Halliburton, and Duke Energy also landing in the top 20. AIDE CEO Paul Cheek, an MIT senior lecturer, says the goal is to give boards an objective peer comparison instead of a survey guess, though the index does not measure whether AI is actually driving financial returns. (Read More)

🫖 TEA For Thought: “Being AI-pilled isn’t just employees using AI. It’s a structural change, a paradigm shift. Surprisingly, several energy companies made the list!”

3. 🐚 The Bedazzled Cyberdeck Rebellion Against Big Tech

A growing community of women on TikTok and Instagram is hand-building “cyberdecks,” small DIY computers (a term coined in William Gibson’s 1984 novel Neuromancer) wrapped in artistic, hyper-feminine shells. Most run on a Raspberry Pi: a seashell mermaid purse that doubles as an e-reader and server monitor, a wood-and-moss deck that plays Game Boy games, a 3D-printed fossil MP3 player, and a Barbie dollhouse that opens into a working mini-computer. Creator CC documents her builds on a blog called Bimbo Tech so other women can follow along even if they do not yet know what RAM is. Blockchain developer Maro Vardanyan weaves pink Raspberry Pis into purses and corsets she calls “macrame motherboards,” a nod to the women textile workers who hand-wove the core-rope memory for NASA’s Apollo Guidance Computer. The movement is framed as a rejection of big tech’s sealed black boxes and a push for radical ownership over the devices people use every day. (Read More)

🫖 TEA For Thought: “We’ll probably see more of these personal cyberdecks doing agentic work for us. It’s the best time to be a builder and creator, and not just a content creator, but a physical, tangible creator!”

4. 🎥 X Bets On Video Reactions To Win Over Creators

On Tuesday, Elon Musk’s X launched “React with Video,” a feature that lets users record a video response to a post instead of reposting or quote-tweeting it. It is pitched as a tool for creators and news influencers who build their brands on commentary and hot takes. X head of product Nikita Bier said “commentary is one of the most important pillars of X,” and that video is sometimes the best way to share a thought. The feature is iOS-only for now, with Android and web coming “soon,” and X says video lets audiences read tone and facial expression that text cannot carry. The launch follows a turbulent stretch for creators: X wound down its Communities feature in April, walked back changes to its revenue-sharing program after backlash, and cut payouts to clickbait accounts. X reported 550 million users as of March 2026, up from 520 million in December, per SpaceX’s S-1 filing. (Read More)

🫖 TEA For Thought: “This is interesting given the X algorithm that went public a couple of weeks ago. It showed that X promotes video more than text posts. People say they prefer to read text, but when everything is text-heavy and everyone’s attention span is shorter, a short video is definitely the catch.”

5. 🔍 Perplexity Turns Search Into Code That Agents Can Write

Perplexity unveiled “Search as Code” (SaC), a new architecture that lets an AI model write Python code to drive the individual parts of a search pipeline rather than calling one fixed, end-to-end search box. It has three layers: the model as a control plane that generates task-specific code, secure sandboxes that run deterministic steps like batching, filtering, and aggregation, and an “Agentic Search SDK” that exposes search as small composable primitives. Intermediate results are saved to a filesystem between turns to avoid burning tokens. Perplexity reports state-of-the-art results on five benchmarks, including 87.1% accuracy on DSQA (versus 73.3% for OpenAI) and 2.5 times the score of rivals on a new benchmark it calls WANDR. In one test that flagged more than 200 high-severity CVEs, SaC hit 100% accuracy while using 85% fewer tokens (42,900 versus 288,700). It is available now through Perplexity Computer and the Agent API. (Read More)

🫖 TEA For Thought: “When I saw this news yesterday, it didn’t register with me. Now I get it: Perplexity is basically doing what AWS does, selling what they’re best at as a service. Like Amazon selling AWS, or selling logistics and shipping as a service, Perplexity is now selling its search backend as a service. How brilliant!”

🛠️ Skill of the Day

The “Definition of Done” Coach: turn a fuzzy goal into a crisp finish line you can actually hit.

You are a pragmatic project coach. I will describe a goal or task that
feels vague or open-ended. Help me define what "done" actually means so
I stop spinning and know when to ship.

Here is my goal:
[DESCRIBE YOUR GOAL, PROJECT, OR TASK]

Do the following:
1. Restate the goal in one plain-language sentence.
2. Define "done" as 3 to 5 concrete, checkable criteria. Each must be
   observable by someone other than me (no "feels good enough").
3. Separate the must-haves for version 1 from the nice-to-haves I can
   cut. Be ruthless.
4. Name the top 2 ways this could fail or drag on, with one guardrail
   for each.
5. Give me a single stop rule: the signal that tells me to ship now
   instead of polishing forever.

Keep it short and specific. If my goal is too big for one pass, propose
the smallest version worth finishing first.

Paste into ChatGPT, Claude, or your tool of choice, drop in any goal, and you get a finish line instead of an endless to-do list.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. 🏦 Paxos Becomes The First Blockchain Firm Cleared To Settle US Stocks

Paxos Securities Settlement Company (PSSC) won full SEC registration to operate as a central securities depository and clearing agency, making it the first blockchain firm authorized to clear and settle traditional US equities onchain. The approval, granted May 29, 2026, caps a path that began with SEC no-action relief in 2019 and a live settlement pilot in February 2020 involving Bank of America, Credit Suisse, and Societe Generale. PSSC can now offer same-day or near-instant settlement, collapsing the traditional multi-day window and freeing capital that institutions keep locked during clearing. The move puts Paxos, led by CEO Charles Cascarilla, in direct competition with the incumbent Depository Trust and Clearing Corporation (DTCC), and clears a key bottleneck for its plan to tokenize real-world assets. Paxos already runs white-label crypto infrastructure for PayPal and Mastercard. (Read More)

🫖 TEA For Thought: “This is huge. Another milestone toward solidifying its foothold in RWA.”

2. 🛍️ One Feed To Track Every Online Store

The Mall, a mobile app that launched in October 2025, is building a single personalized feed that tracks products, sales, and drops across more than 10,000 fashion brands. It uses web scraping to pull entire retail catalogs and large language models to categorize products, then lets users follow a brand by pasting an Instagram or TikTok handle and pushes alerts when items go on sale, restock, or drop. Founders Sreya Halder (a Stanford computer science graduate) and Ellie Konsker (previously at Tom Ford and Karla Otto) met through a female founders circle in Los Angeles. The app is free for shoppers, and the company plans to monetize through a B2B analytics product launching summer 2026 that sells brands aggregated, anonymized shopping behavior and competitor assortment data, followed later by ads and brand subscriptions. The Mall is invite-only after a 4,500-person beta, with wide release expected by the end of summer. (Read More)

🫖 TEA For Thought: “This product has definitely figured out its PMF. Pretty inspiring, too, the way they handle monetization.”

3. ⚙️ Huawei Thanks Washington For Supercharging China’s Chip Industry

Huawei rotating chairman Xu Zhijun publicly thanked the United States for its chip export controls, arguing the restrictions forced China to build its own semiconductor stack faster than it otherwise would have. “If the United States hadn’t forced our country, our companies, and our industry, we wouldn’t have done something like this. But we are also grateful to the US for enabling our country’s semiconductor industry chain to truly grow,” he said. The remarks came as Huawei unveiled new chip techniques it calls the Tau Scaling Law and LogicFolding, which aim to boost performance by shortening the distance signals travel rather than only shrinking transistors. Xu said export bans pushed Chinese firms to invest urgently in domestic design, packaging, software tools, and manufacturing, with rising revenue funding further R&D. One marker of the shift: Nvidia’s share of China’s AI chip market has fallen from about 95% to near zero since the bans took hold. (Read More)

🫖 TEA For Thought: “On the one hand, the CCP threatens the US, all sassy and spiky. On the other, it keeps buying up the countries around it to sneak in new chips.”

4. 🚀 Blue Origin’s Rocket Explodes On The Pad, And NASA Feels It

A Blue Origin New Glenn rocket exploded during an evening static fire test on May 28, 2026, at Launch Complex 36 at Cape Canaveral, destroying the vehicle and severely damaging the company’s only New Glenn pad. The test was the final milestone before a launch targeted for June 4. The blast appeared to start at or near the base of the first stage as its seven methane-burning BE-4 engines ignited, though the cause, possibly a ruptured fuel line or a ground-system fault, was not immediately known. The setback ripples straight into NASA’s Moon plans, because Blue Origin’s Blue Moon landers fly on New Glenn. The Blue Moon Mark 1, slated for fall 2026 to deliver building blocks for a lunar base, is now unlikely to launch before late 2027. With Blue Moon grounded indefinitely, NASA grows even more dependent on SpaceX’s Starship for the crewed Artemis III and IV missions. (Read More)

🫖 TEA For Thought: “Rocket science is hard for a reason. Sad for New Glenn, but it further shows how capable SpaceX is as a private, almost public, company, achieving so much under someone who also runs Tesla, X, The Boring Company, and Neuralink...”

5. ⚡ AI Is The New Electricity, And The Org Chart Will Follow

In a widely shared essay, consultant Ted Neward argues that AI will follow the same arc as the internet and electricity: a technology that starts as a C-suite obsession and ends as invisible infrastructure. He notes that companies once appointed “Chief Internet Officers,” much as an earlier era briefly had electricity executives, and that neither role survived once the technology became a utility rather than a novelty. Neward expects the current AI bubble to deflate, but believes firms that treat AI as a means to an end, woven through their operations, will outlast those that treat it as the product itself. He contends frontier models lack a durable moat as open-source and locally trained alternatives erode exclusivity and compute leads prove temporary. His blunt forecast: diversified giants like Apple, Microsoft, Meta, and Google absorb the shock, while companies defined entirely by AI, including OpenAI and Anthropic, are the most exposed, with chipmaker Nvidia somewhere in between. (Read More)

🫖 TEA For Thought: “AI is as revolutionary as the internet and electricity. Intelligence itself will become the commodity, and every business will have to evolve around it. It’s a huge wave that no one can stop.”

🛠️ Skill of the Day

The Two-Sided Briefing: turn any article or hot topic into a balanced brief you can actually reason with.

You are a sharp, even-handed analyst. I will give you an article, a
claim, or a topic. Help me understand it well enough to form my own
view, without you taking a side.

Here is the material:
[PASTE ARTICLE, LINK SUMMARY, OR TOPIC]

Do the following:
1. Plain-language summary: explain what is going on in 5 sentences, as
   if to a smart friend outside the field. Define any jargon inline.
2. The strongest case FOR (the bull case): the best 3 arguments a
   thoughtful supporter would make.
3. The strongest case AGAINST (the bear case): the best 3 arguments a
   thoughtful skeptic would make.
4. What each side must believe: the one key assumption under each case
   that, if it turns out wrong, sinks that side.
5. What to watch next: 2 or 3 concrete signals that would tell me which
   side is winning over the next year.

Be specific and fair. Do not hedge into mush. If the material is thin,
tell me what is missing.

Paste into ChatGPT, Claude, or your tool of choice, drop in any article or topic, and you get a brief you can think with instead of just a summary.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

1. 🗺️ The Top AI County In America Is A College Town

Microsoft’s AI Economy Institute mapped US AI adoption for Q1 2026, and the geography defies expectations. The top AI-using county is not in Silicon Valley. It is Williamsburg, Virginia (home to William & Mary) at 73.7%, followed by Harrisonburg, Virginia (James Madison) at 67.9%, Madison, Idaho (BYU-Idaho) at 67.7%, Brazos County, Texas (Texas A&M) at 64.5%, and Story County, Iowa (Iowa State) at 64.2%. College towns are driving the trend: counties with more 18-to-24-year-olds average 28.6% usage versus 20.3% elsewhere. At the state level, DC leads at 40.6%, then Maryland (36.5%) and Utah (35.9%), with Texas (35.4%) edging out California (34.1%) and New York (32.9%). Nationally, 31.3% of working-age Americans now use AI, lifting the US from 24th to 21st globally. The urban-rural gap stays wide: 32.9% in metro counties versus 16.2% in rural areas. (Read More)

🫖 TEA For Thought: “It’s not as what you’d image”

2. 🪙 Tokenized Assets Proved The Concept. Now Comes The Hard Part.

a16z crypto published 7 charts showing the tokenized real-world asset (RWA) market crossed $30 billion and now sits near $34 billion, excluding stablecoins, a 10x jump from under $3 billion in mid-2024. The GENIUS Act’s clearer US stablecoin rules plus maturing institutional infrastructure drove the surge. Tokenized US Treasuries and commodities make up roughly two-thirds of the market: bonds lead at $15.2 billion (BlackRock and Franklin Templeton dominate), commodities sit near $5.1 billion (almost all gold via Tether’s XAUT and Paxos’s PAXG), and tokenized stocks total about $1.5 billion. Ethereum hosts $15.7 billion, ahead of BNB Chain ($4B) and Solana ($2.2B). The catch: Pantera Capital’s token presence index ranks more than three-quarters of assets in the lowest tier, meaning most are “digital receipts” rather than truly composable onchain assets. Only about 5% of bonds are deployed in DeFi. The hard part is making tokenized assets genuinely programmable, not just digitized. (Read More)

🫖 TEA For Thought: “This is just the beginning.”

3. 💼 The Agentic CFO In Your Pocket

In a CoinDesk opinion piece, Joseph Chalom argues that AI agents powered by stablecoins, tokenized assets, and DeFi will hand retail savers the kind of treasury management once reserved for institutions. The scale of the gap is stark: American households hold $6 trillion in checking and nearly $15 trillion in savings, yet US retail savers forgo at least $180 billion in interest every year. An autonomous agent could monitor real-time cash flows, run securities lending, and vote shares automatically. The infrastructure is arriving fast: the stablecoin market is projected to grow from $330 billion to $3 trillion by 2030, the tokenized-asset industry could reach $100 trillion by decade’s end, and the open X402 payments protocol has already logged 167 million agent-to-agent transactions this year. Standards from Visa, Mastercard, Google, and Stripe (which processed $1.9 trillion in volume last year) are converging on agent payments. (Read More)

🫖 TEA For Thought: “AI is the biggest equalizer of all times.”

4. 🧩 Where AI’s Real Product-Market Fit Lives

Simon Willison argues OpenAI and Anthropic found genuine product-market fit not in consumer chatbots but in enterprise coding agents like Claude Code and Codex. The economics tell the story: he personally burns around $2,180 a month in API tokens ($1,199.79 to Anthropic, $980.37 to OpenAI) while paying just $200 in subscriptions, so both firms moved enterprise plans to pay-per-token pricing in April 2026. Anthropic’s revenue reached $10.9 billion in the second quarter, and SpaceX agreed to pay it $1.25 billion per month through 2029 for compute. By contrast, ChatGPT has 900M-plus weekly users but only 5.6% pay. Willison also calls the viral “company blew its AI budget” stories overblown, including claims about Uber’s CTO maxing out a full-year budget. His thesis: coding agents win because they consume far more tokens while serving highly paid knowledge workers. (Read More)

🫖 TEA For Thought: “Enterprise is where the money is. In the agentic era, B2B is where money lies, not so much B2C, as the amount of money an individual user is willing to pay is too little compared to the cost it generates.”

5. 📊 The 1% Of Developers Who Eat All The Tokens

Cursor’s Spring 2026 Developer Habits report reveals that AI usage among developers is extraordinarily concentrated. P99 developers produce 46x more AI-generated lines than the median user and merge 15x more pull requests. The Gini coefficients are striking: 0.77 for AI lines, 0.75 for spend, 0.72 for tokens, with the top 5% accounting for the vast majority of all AI activity. Output is accelerating across the board: lines added per developer per week climbed from 3.6K in January 2025 to 8.6K in May 2026, “mega PRs” of 1,000-plus lines grew from 8% to 13.8% of merges, and AI-generated code survival rose from 76.6% to 80.6%. Autonomy is surging too: AI changes accepted without manual review jumped from 7% in January to 36.3% in May 2026. Model costs vary 9x per request, from $0.18 (Composer 2.5) to $1.57 (Opus 4.7). (Read More)

🫖 TEA For Thought: “AI usage is highly concentrated, with a small share of developers accounting for a large share of AI lines, spend, and token consumption.”

🛠️ Skill of the Day

The Cost-Per-Outcome Auditor: turn any “should we buy/build this AI tool?” question into a clear one-page decision memo with real numbers.

You are a pragmatic operations analyst. I am deciding whether to adopt
[TOOL OR APPROACH] for [TASK / TEAM]. Help me decide based on cost per
outcome, not hype.

Here is my situation:
- What we do today: [CURRENT PROCESS]
- What the tool/approach promises: [PASTE PITCH OR DESCRIPTION]
- Rough usage: [HOW OFTEN / HOW MANY PEOPLE / VOLUME]
- Budget reality: [WHAT WE PAY NOW / WHAT WE CAN SPEND]

Produce a one-page memo with these sections:
1. Unit of value: define the single outcome that matters here.
2. Cost per outcome, three scenarios: low, expected, and runaway usage.
   Show the math and name every assumption you make.
3. The cheaper alternative I might be ignoring (including "do nothing").
4. The breakeven point: at what volume does this win or lose?
5. Verdict in one sentence, plus the top risk that would flip it.

Ask me up to three questions before you start only if a number you need
is missing. Keep it concrete. No buzzwords.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own, and let it pressure-test the spend before you commit.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow!

Subscribe now

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

Phase One: The Cost Reduction Era

Most companies today are using AI to automate existing workflows and cut costs. We ask AI to use our computers, submit forms, and navigate graphical user interfaces that were originally designed for human eyes. This is the dominant use case today. Value creation through AI is still rare, mostly concentrated at the foundation model layer where companies like OpenAI and Anthropic are finding strong product-market fit by providing the core tools.

This phase is real and necessary, but it is fundamentally temporary. The old paradigm is the launchpad, not the destination. Eventually, the focus will shift from merely reducing the cost of old processes to generating entirely new forms of value.

The New Paradigm: Build for the Future, Not the Ceiling

The rules of the game are changing. GUIs built for humans are inefficient and wasteful for AI. Agent-facing interfaces, new interaction models, and entirely new problem categories are emerging.

We see this shift happening in real time. He Xiaopeng scrapped billions in legacy architecture to bet entirely on end-to-end AI for autonomous driving. The founders of Manus waited for the right model from Anthropic before launching their product, knowing their vision required future capabilities. As Garry Tan, the CEO of Y Combinator, points out, the only true failure mode is staying stagnant. The wave is moving, and it will not wait for those who are fixated on the present.

Leaders do not optimize for today’s constraints; they aim at tomorrow’s capabilities. Innovation is simply belief aimed at the future. If everyone focuses only on what works right now, true innovation becomes impossible.

AI as the Great Equalizer (and Its Shadow)

AI is rapidly lowering the barrier to building. We are entering an era where business logic matters far more than technical fundamentals. You do not need to know how every specific function of a coding language works if you understand the architecture, the workflow, and the problem you are trying to solve. Anyone who asks the right questions can build.

But this equalizer cuts both ways. While it empowers many, it also widens the gap between those who think clearly and those who do not. Those who are curious and driven will compound their advantage. It is the Matthew Effect in action: “For whoever has will be given more, and they will have an abundance.“

Conversely, those who outsource their thinking risk losing the very agency that makes them human. If we rely entirely on AI without understanding our own principles, we risk becoming dependent on the systems we built. The equalizer only works if you stay awake to it.

Which Future Are You Going Back To?

The future is not predetermined. It is whatever you prioritize building toward.

God has given every person creativity, curiosity, and the capacity to ask profound questions. AI, at its best, is a tool that liberates people to use those gifts more fully. When we spend less time on busy work, we have more time to create, to learn, and to seek the true meaning of life.

The wave is moving. The question is not whether to engage with it, but which future you are building toward, and whether that future is worthy of the gifts you have been given.

Which future are you excited to go back to?

Till next time, Cheers!

Subscribe now

Previous column articles can be found here:

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

1. 🔓 12 AI Prompts That Quietly Leak Enterprise Data

ChatGPT triggered 410 million data loss prevention (DLP) policy violations in a single year, a 99.3% jump year over year, per Zscaler’s ThreatLabz 2026 AI Security Report. The article maps 12 everyday work scenarios where company data slips out: pasting vendor contracts for summaries, polishing HR reviews with names and pay, uploading resumes, cleaning CRM records, drafting sales emails from internal account notes, debugging proprietary code in public assistants, and even pasting live API tokens during troubleshooting. The fix is calibrated, not a blanket ban: allow and log non-sensitive data, warn before low-severity submissions, hard-block credentials, source code, and regulated PII, plus automated redaction, clipboard isolation, and AI audit trails. (Read More)

🫖 TEA For Thought: “Bring your own AI to work is just something every employee does nowadays.”

2. 💸 Companies Are Spending Billions to Train Workers for AI. Most of It Will Fail.

Mercer’s 2026 Global Talent Trends study found 63% of executives see AI work redesign as their highest-return investment, yet only 32% say their workforce is actually ready, and 98% plan org redesigns within two years. The article argues most training spend fails because companies skip the diagnostic step: they fund the “how” (training content) without measuring the “who” (which employees have the behavioral profile to adapt). Resumes show past experience, performance reviews measure yesterday’s job, and AI fluency surveys only gauge familiarity with tools people barely use. The proposed fix reverses the sequence: measure top performers’ behavior first, redesign roles on evidence, then target training to the gaps. One supply distributor that built a behavioral model before training saw 20% more net sales, about $89,000 extra revenue per employee. (Read More)

🫖 TEA For Thought: “AI teaming is not just another compliance training. The bigger the enterprise is, the harder it is to become AI-pilled. If that is the case, how do they pivot? It is not just a simple change, it is a transformation inside out.”

3. 📊 ITBench: How Much Can AI Agents Actually Finish on Their Own

Artificial Analysis built ITBench-AA, an independent implementation of IBM’s ITBench, to measure how well AI agents handle real IT automation across Site Reliability Engineering (SRE), security and compliance (CISO), and Financial Operations (FinOps). For SRE, agents analyzed 59 Kubernetes incident snapshots and had to find the root-cause entity, such as a misconfigured feature flag or a missing container image. On the SRE precision metric, Claude Opus 4.7 led at 46.7%, GPT-5.5 followed at 45.8%, and Qwen3.7 Max hit 42.5%. The harder headline: across full task resolution, the best agent resolved only 13.8% of SRE scenarios, 25.2% of CISO scenarios, and 0% of FinOps. The benchmark measures finished work, not partial credit. (Read More)

🫖 TEA For Thought: “This benchmark makes sense, as what agents can actually get done independently is the goal at the end of the day.”

4. 🪪 Uber Solves the AI Agent Identity Crisis

Uber detailed how it gave AI agents real identities. The problem: traditional identity systems built for humans and service accounts cannot express “agency,” the idea that an agent acts on behalf of a user. When one agent hands a task to another, downstream systems only see a generic service account, and the originating user plus intermediate agents get dropped across hops, breaking the audit trail. Uber’s fix combines an Agent Registry (source of truth for agent-to-workload mapping), a Security Token Service that mints short-lived, single-hop JWT tokens carrying an embedded “actor chain” (user to agent to agent to system), an AI Agent Mesh for communication, an MCP Gateway for policy enforcement, and SPIRE-issued cryptographic SVIDs as the identity foundation. Deployed across thousands of internal agents, token exchange holds P99 latency under 40 milliseconds. (Read More)

🫖 TEA For Thought: “Agent identity is one of the combo points where AI and blockchain meet!”

5. 📈 Just Like Gold and Oil, We Will Soon Trade AI Token Futures

A derivatives market for AI compute is forming. The Shanghai Futures Exchange is designing a market for AI tokens, while CME Group and Intercontinental Exchange (ICE, owner of NYSE) are launching GPU compute futures. The pitch: treat AI tokens like a commodity, similar to electricity or bandwidth, so businesses and data center operators can hedge against swinging compute costs. The prices are real and volatile. OpenAI charges $5 per million input tokens and $30 per million output tokens for GPT-5.5 API access, H100 GPU rentals run $1.40 to $4.27 per hour, and H200s run $2.34 to $5 per hour. Cloud providers including Amazon Bedrock, Google Cloud, and Oracle increasingly price per token, setting the stage for tokens to trade like any other raw material. (Read More)

🫖 TEA For Thought: “Just like gold and oil, we’ll soon be able to trade AI token futures.”

🛠️ Skill of the Day

Sanitize Before You Paste: a quick gatekeeper that flags sensitive data in anything you are about to send to an AI tool, before it leaves your hands.

You are my data privacy gatekeeper. I am about to paste the text below into a public AI tool. Before I do, scan it and tell me what I should remove or mask.

Find and list any: full names, emails, phone numbers, home or office addresses, customer or employee records, salary or financial figures, contract terms, API keys, passwords, tokens, internal project names, unreleased product details, or regulated data (health, legal, financial).

Then give me:
1. A risk rating: SAFE, CAUTION, or DO NOT PASTE.
2. A bullet list of every sensitive item you found and where it appears.
3. A cleaned, rewritten version with each sensitive item replaced by a neutral placeholder like [CLIENT NAME] or [DOLLAR AMOUNT], so the AI can still help me with the task.

Do not summarize or answer the underlying task yet. Just sanitize.

TEXT TO CHECK:
[PASTE YOUR TEXT HERE]

Paste into ChatGPT, Claude, or your tool of choice. Run it on anything work-related before you share it, then use the cleaned version for your actual request.

TEAHEE Moment

Stay sharp, stay informed. See you Monday.

Subscribe now

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

1. 🧠 Language models may need “sleep”

Transformer-based LLMs struggle as context grows because attention cost scales poorly. A new approach proposes a “sleep-like” consolidation step where the model periodically reprocesses recent context offline, updates persistent fast weights via a learned local rule, clears the KV cache, then resumes inference with the consolidated memory. Tests on synthetic reasoning tasks and a long-horizon math reasoning setting show performance improves as “sleep duration” (number of offline passes) increases, with the biggest gains on deeper reasoning cases. (Read More)

🫖 TEA For Thought: “The point of dreaming.”

2. 🔍 Nobody owns “correct”, teams need “verifiable”

Harvey AI rebuilt a document review system that was already working because it was “insufficiently verifiable”, meaning it did not show evidence and reasoning at a per-statement level. The piece argues that “insufficiently verifiable” is a failure mode separate from “incorrect”, and that many orgs optimize for benchmark scores while missing whether outputs can be checked, trusted, and safely acted on. It highlights how AI evaluation can be gamed by non-answers that look good numerically, and why verification needs deterministic guardrails for scope and permissions, not just LLM judging. (Read More)

🫖 TEA For Thought: “Insufficiently verifiable is a failure mode distinct from incorrect.”

3. 🤖 CEOs and “AI psychosis” (happy-path delusions)

A TechCrunch piece argues some CEOs are overestimating what agents can do because they see AI demos on the “happy path” and are far from the last-mile work: reviewing code, finding bugs, catching hallucinated libraries, and dealing with edge cases. It cites Box CEO Aaron Levie’s framing that executives are uniquely prone to “AI psychosis”, plus notes 2026 layoffs have already nearly matched 2025 levels, with many companies pointing to AI as a rationale. It also points to research suggesting measured productivity gains often lag perceived gains, and that organizational bottlenecks can shift to executives when AI increases output volume. (Read More)

🫖 TEA For Thought: “But the thing is, a CEO is not just focusing on here and now, but also what AI could do later down the road. A lot of things that AI does now can be buggy, but with the improvement of models, which happens so fast and so frequently, they will not have to be the blocker anymore.”

4. 🔐 MFA resets and token theft are the new front door

An emerging dominant pattern in financial services intrusions is not stealing passwords, but getting MFA reset (often through social engineering) and capturing OAuth tokens that grant persistent access. The piece cites multiple reports: CrowdStrike describing vishing over Microsoft Teams to convince employees to reset MFA and register attacker devices, the FBI warning about token theft via device code flow, and Verizon DBIR data showing vulnerability exploitation and identity-based paths shifting. The takeaway is MFA protects password-based login, but attackers are increasingly bypassing it through resets, tokens, and workflow abuse. (Read More)

🫖 TEA For Thought: “This is even scarier when we give agents so many accesses via OAuth.”

5. 📈 DeepSeek tops OpenRouter usage rankings

OpenRouter usage data reportedly shows DeepSeek-V4-Flash leading global model usage with a weekly invocation volume of 3.43 trillion tokens, while total model usage reached 28.9 trillion tokens from May 18 to May 24, up 7.4% week-over-week. The article attributes growth to broader AI agent adoption and workloads like code generation, long-document processing, and enterprise retrieval that drive very large token consumption. It also reports Chinese model usage at 9.223 trillion tokens weekly versus US model usage at 4.93 trillion, with China leading for multiple consecutive weeks. (Read More)

🫖 TEA For Thought: “This will only go up exponentially.”

🛠️ Skill of the Day

Reality Check Prompt: sanity-test an “AI can do this” claim before you commit budget, headcount, or reputation.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

Twitter/X

1. 🔄 NEAR Rallies 15% In A Day As Cross-Chain Intents Cross $19B

NEAR token climbed 15% in 24 hours to $2.8, extending a month-long run of roughly 90% gains. The rally is driven by NEAR Intents, a cross-chain transaction layer that has processed over $19 billion in cumulative volume and generated about $32 million in fees by letting users request outcomes (like swapping USDC on Ethereum for SOL on Solana) while third-party solvers handle execution. BitMEX co-founder Arthur Hayes called NEAR, Hyperliquid’s HYPE, and ZEC crypto’s “holy trinity.” Institutional flow is showing up too: the Bitwise NEAR Staking ETP in Europe now holds about $40 million in assets, with $7 million in inflows in a single week. A June network upgrade will introduce dynamic resharding to auto-split shards under load. NEAR still trades far below its 2022 peak near $20. (Read More)

🫖 TEA For Thought: “Let’s not forget that NEAR started off as an AI company and then switched to blockchain.”

2. 🤖 Human Archive Raises $8.2M To Pay Indian Workers $1/Hr To Train Robots

Human Archive, founded by UC Berkeley students Samay Maini, Rushil Agarwal, and Shloke Patel along with Stanford’s Raj Patel (CEO), raised $8.2 million from Wing Venture Capital, NVP Capital, Y Combinator, and angels at OpenAI, Nvidia, Google, and Meta. The startup pays Indian gig workers roughly $1 per hour to wear camera-equipped caps, tactile gloves, motion capture suits, and wrist cameras while doing everyday service work. It has deployed over 1,000 active headsets across India’s home services, hotel, and restaurant sectors after being rejected by Urban Company and Pronto. Wing VC says the rig captures synchronized RGB-D video, force feedback, and full-body motion at scale. Consumers get discounted services in return for consenting to recording, with faces blurred under India’s Digital Personal Data Protection Act. Expansion into Southeast Asia and the U.S. is underway. (Read More)

🫖 TEA For Thought: “The era of downloading the internet is over. The era of uploading humans is here.”

3. 🛡️ Cybersecurity Hiring Up 11% As Anthropic And OpenAI Ship Offensive Models

Cybersecurity job postings in Q1 rose 11% year-over-year per Glassdoor, while executive search firm Heidrick & Struggles reports a “five-, maybe sevenfold” jump in demand since last fall. Roles that used to appear once a year are now requested weekly by Fortune 100 companies, and search firms are turning clients away. Anthropic recently released its Mythos model, which can identify and exploit vulnerabilities in critical infrastructure like power grids and financial systems. OpenAI followed with GPT-5.4-Cyber. LinkedIn CISO Lea Kissner put it plainly: “We’re going to need people to deal with the bug-pocalypse.” Security engineer Brian Gaudenti is among those adapting by learning AI tooling, since defenders increasingly use the same models to generate code (and sometimes the same vulnerabilities), which paradoxically drives further hiring. (Read More)

🫖 TEA For Thought: “AI is way better at offense than defense. For now, humans paired with a super AI might be able to defend. But soon it will have to be AI against AI. Swarms of AI against swarms of AI.”

4. 🐢 Engineer Argues AI Lets You Ship Better Code, Slower

Nolan Lawson argues the prevailing “AI = fast, sloppy code” narrative misses a quieter use case: using LLMs to ship higher-quality code at a more measured pace. His workflow runs Claude (Opus 4.7 on extended thinking), GPT 5.5 Codex, and Cursor Bugbot in parallel as code reviewers, then consolidates their findings across critical, high, medium, and low severity tiers to filter false positives via cross-model validation. Fixes start with critical and high issues, skip low-ROI items, and abandon fundamentally flawed approaches. Lawson documents heavily with Markdown and Mermaid diagrams so future readers can follow the PR. He warns the method “finds so many bugs that you’ll be bored senseless if you try to tackle them all,” because it surfaces pre-existing code rot, turning feature work into deeper architectural side-quests that compound long-term codebase understanding. (Read More)

🫖 TEA For Thought: “Make better better by aiming at better.”

5. 🌐 Wix Cuts 1,000 Jobs As AI Threatens The Website-Builder Business Model

Wix is cutting roughly 1,000 jobs, about 20% of its 5,277-person workforce (over 60% Israel-based at the Glilot Junction campus), as investors worry website builders become irrelevant when anyone can prompt a site in five minutes. The stock has dropped nearly 50% since early 2026, dragging the valuation to about $2 billion. Q1 2026 posted a $57.5 million net loss on $541 million revenue (up 14%), with cash flow down 21% to $112 million and operating expenses up 50% to $423 million (now 35% of revenue, up from 21%). CEO Avishai Abrahami is pivoting via the Base44 acquisition (natural-language AI programming, $150 million ARR by May 2026) while Wix simultaneously builds its own AI model, a move that is pushing compute costs higher even as it tries to define a post-prompt-first business. (Read More)

🫖 TEA For Thought: “This is tough. Business models like this are doomed in the AI era when any website can be generated with one prompt in 5 minutes. The real question is what could Wix have pivoted to before this happened? I’d pivot toward agentifying the websites: make websites agent-facing and agent-friendly. AEO is still a huge market, and current websites are not ready for agents yet.”

🛠️ Skill of the Day

Pre-Mortem Planner: stress-test any plan or decision by imagining it has already failed, then work backwards to find what killed it.

You are my pre-mortem strategist. I am about to commit to the plan or decision below. Before I commit, I want you to imagine it is now six months in the future and this plan has clearly failed. Your job is to figure out why, in concrete terms, and help me fix it before I start.

Plan or decision: [PASTE YOUR PLAN, DECISION, OR PROJECT BRIEF HERE]

Do this in five steps:
1. List the top 7 most likely failure modes. Be specific (not "execution risk"). Include at least two failures driven by people or incentives, not just tech.
2. For each failure, rate likelihood (1 to 10) and severity (1 to 10), then sort by likelihood times severity.
3. For the top 3 risks, write a one-sentence "early warning signal" I could detect in the first 30 days.
4. Propose one concrete change to the plan that defuses each top risk without killing the upside.
5. End with a one-paragraph rewritten plan that integrates your fixes. Plain language, no jargon.

Be blunt. If the whole plan is fragile, say so and propose a simpler alternative.

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

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

Subscribe now

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

Overnight, my AI agents had been running in parallel. By the time I opened my laptop, eight different projects were waiting for me. They had all reached a point where they needed a strategic decision to move forward.

I looked at the first one, loaded the context into my brain, and made the call. I did the same for the second. By the third, I was completely exhausted. I simply did not have the capacity to review the rest of the work and give them the right instructions. And these were important decisions—if I didn’t read through them carefully and give clear direction, the agents would just waste compute doing something completely wrong.

The bottleneck wasn’t the AI. The bottleneck was me. More specifically, it was my capacity for deep, high-quality judgment.

This experience got me thinking about context management, and how the way AI systems work is actually a perfect mirror for how human beings operate.

The Mirror

Right now, AI engineering is obsessed with compute allocation. Compute means tokens, and tokens are what generate value. Even companies that promote “token maximization” do it because they want their engineers to find ways to create value. But compute is finite. How do you allocate it? What is your North Star? It comes down to priorities. As Elon Musk has pointed out, if you just focus on removing the one main blocker every week, everything else starts to resolve itself.

For humans, compute is attention. There is a famous AI paper called “Attention Is All You Need”, and for human beings, that is literally true. Attention is all we have. It takes immense energy and thinking capacity to focus on the things that matter. When you constantly have to make strategic decisions, you run into decision fatigue. Your capacity to think strategically is limited.

Once priorities are set, the next challenge is executing the task. In AI, this is called context engineering—managing what goes into the context window. The context window is just like our working memory. If you load all the learnings, all the patterns, and the entire big picture into the window at the beginning, there is no space left to actually process the specific task at hand.

It is the same for us. If you have five projects loaded in your working memory, along with the financial worries and the grand vision, you will freeze. There is a Chinese saying: your eyes are looking too high up, and you are not able to go low enough to pick up the low-hanging fruit. If you only look at the goal, you will never take the steps toward it.

This brings us to the third concept: harness engineering. A harness is the scaffolding you put around an AI agent to help it work better—knowing what to remember, what to forget, and how to self-repair.

But what is the harness for us? How do we manage our own capacity when we have multiple projects running simultaneously?

Building Your Own Harness

I am still figuring this out, but here is what I am learning about managing my own attention so I can actually generate value instead of just burning out:

1. Batching by cognitive mode, not by project. Strategic decisions require slow, deliberate, high-context thinking. Instead of context-switching between projects throughout the day, I am realizing the value of designating one block of time—maybe the first 90 minutes of the day—as a decision window. Outside of that window, I am in execution mode.

2. Pre-loading context the night before. Reviewing eight projects in the morning is exhausting because you are doing two things at once: loading the context and making the decision. If you spend 15 minutes the evening before just reading the status updates without making any decisions, your working memory is primed for the next morning. The decision becomes much cheaper.

3. The “one project per day” depth rule. For each project, designate one day a week where it gets your deepest attention. The other days, it only gets reactive responses. You cannot go deep on five things in one morning, but you can spread that depth across the week.

4. Letting the context breathe. Sometimes, when you hit decision fatigue, the best thing to do is write down the one question you need to answer, and sleep on it. Your brain consolidates information when you rest. The answer often surfaces faster the next day.

5. Externalizing the meta-context. In a multi-agent system, an orchestrator holds the big picture so individual agents don’t have to. You can do this for yourself: keep a single document updated weekly with your priorities and the status of each project. Write it down so your brain doesn’t have to hold it all in working memory.

The Ultimate “Why”

These frameworks are borrowed directly from AI architecture. We are applying AI engineering back to ourselves. But all of this attention management leads to a much bigger question.

AI is the means, not the ends. We cannot mix up means with ends.

If AI is freeing us from busy work, repetitive tasks, and execution, what exactly are we freeing our attention for?

In the past, great thinkers had the leisure to ponder the meaning of life because their basic needs were taken care of. AI has the potential to do that for us at scale. If we don’t have to do the busy work, we have the time and energy to create, and to ask the ultimate questions. What is the meaning of life? Why do we live?

For me, the ultimate answer is God. As Viktor Frankl noted in Man’s Search for Meaning, the search for meaning is often the answer itself. I want to use the time AI gives me to understand God better, and to bring love to others.

As Nietzsche said, “He who has a why to live for can bear almost any how.” AI is increasingly figuring out the how. That makes our why the most important thing we possess.

The Warning

But there is a flip side to this freedom.

What will people actually do with their free time? History shows that rulers often keep people busy so they never ask the big questions. If we are not careful, AI could become a new kind of control.

If we outsource all our thinking, our memory, and our work to AI—and if that AI is controlled by a few massive corporations—we become digital colonies. We become slaves without even realizing it.

When you outsource your judgment to AI, you stop training your own thinking muscles. If you don’t criticize the work, you lose the ability to differentiate good from bad. You lose your taste. And once you lose your taste, you lose your agency.

Attention is your most valuable resource. Jordan Peterson once noted that attention is basically worship—whatever you pay the most attention to becomes your god.

In an era where knowledge is democratized and execution is cheap, your attention and your judgment are all you have left. Build a harness to protect them. Figure out your why. Because if you don’t decide what to pay attention to, someone else will decide for you.

Till next time, Cheers!

Subscribe now

Previous column articles can be found here:

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