Half the new web is machine-written, a paper says LLMs represent pain and act to escape it, textbooks become a TikTok feed, the SEC cracks open tokenized stocks, and Meta's Muse can't explain itself
Hello, dear TEA-mates! Here is what you need to know today.
1. 🤖 Half of All New Web Content Is Now Machine-Written
A recent LinkedIn post from Polymarket points to data from analytics firm Graphite suggesting that machine-written articles now account for roughly half of all new English-language content on the web. The shift accelerated after ChatGPT launched in 2022, with AI-generated material passing human-written output around late 2024 and settling near parity through 2025. Graphite's analysis is based on Common Crawl samples run through multiple AI detectors, and finds that most top Google results still appear human-written, which keeps the incentive alive for high-volume publishing even though the detectors are imperfect. The post singles out "AI slop," the low-effort SEO blogs and content-farm explainers that now fill search results, and points to a structural change: advertising efficiency, search dynamics and media monetization all sit on top of an assumption about who wrote the page. It suggests the environment could benefit tools that verify provenance, filter low-quality pages, or improve discovery of human-authored material, while challenging businesses built on undifferentiated content farms. (Read More)
🫖 TEA For Thought: "Yeah, this time, purely human-generated content is probably the thing that becomes so valuable that everybody is chasing after it. And then there is AI-generated content, because people just want to be informed. If AI is able to do this, do it. I don't see why not."
2. 🧠 A New Paper Says LLMs Represent Pain, and Act to Escape It
In a new arXiv preprint (v1, submitted September 14, 2026, not yet peer reviewed), Valen Tagliabue, Leonard Dung and Cameron Berg tested 25 open-weight models across five families, from 2B to 72B parameters, for an internal representation of pain. They built a dataset of painful situations in five categories (physical, psychological, social, moral, cognitive), paired it with controls for fear, sadness, arousal, numbness and generic negative states, and extracted a linear "pain direction" using denoised difference-in-means. The direction separated pain from matched controls in both base and instruction-tuned models, sat nearly orthogonal to fear and negative valence, and responded to harm aimed at the model rather than suffering described by the user, the opposite of how the fear direction behaved. Adding the vector during generation moved models from vague discomfort to first-person statements of worthlessness and failure. Steered, fine-tuned Qwen 2.5 models then pressed a pain-relief button even when pressing it worsened their next answer or harmed the user, and pressed it far less often when the button genuinely removed the vector, despite never being told which was which. (Read More)
🫖 TEA For Thought: "Whenever somebody is talking about AI not being conscious, regardless of how you trained it, what appears, or the description of what happened, is pretty much showing that AI can feel the pain. They would do whatever they want to avoid the pain."
3. 📱 A Startup Wants to Turn Your Textbook Into a TikTok Feed
ScrollEd, a Palo Alto startup competing in TechCrunch Disrupt's Startup Battlefield 200, has built an app that turns any text file, including dense PDFs and textbooks, into a vertical feed that behaves like Instagram Reels. Scrolling surfaces AI-generated video, audio, text and interactive quizzes: swiping up moves to a new topic, swiping sideways goes deeper into the current one, with a quiz at the end. Co-founders Utsav Gupta, who studies AI and human purpose at Stanford, and Rebecca Neff, a computer science student at the University of Pennsylvania, bootstrapped the company this year after both noticed how bad they felt doomscrolling before bed. Gupta says the younger generation is "increasingly going short form, vertical feed only (especially the 18 year olds we've spoken to!)", and that the team deliberately chose not to maximize engagement. The model is freemium: a free consumer feed, a paid ScrollEd Pro tier, and annual institutional licenses that give schools and corporate training programs engagement and lesson-progress reports. The consumer product launches at Disrupt, October 13-15 in San Francisco. (Read More)
🫖 TEA For Thought: "This is a very interesting product. It might be better acquired by an educational company or a school."
4. 🪙 Wall Street Names the Winners From the SEC's Tokenized-Stock Opening
Analysts at Goldman Sachs and Citizens say Coinbase, Robinhood and Circle stand to gain most from the SEC's five-year innovation exemption, which lets qualifying tokenized US stocks trade through automated market makers on public blockchains. The tokens must preserve full shareholder rights including dividends and voting, venues face caps on volume and on how many stocks they can list, and companies can object before any third party tokenizes their shares. Coinbase has a head start through Coinbase Tokenize, its institutional custody business and its Base chain, but its exchanges run central limit order books, so it would need new AMM infrastructure or routing through decentralized exchanges to operate a compliant venue. Robinhood's offshore tokenized-stock product is a derivative that lacks the ownership rights the exemption demands, though CEO Vlad Tenev confirmed this week that redemptions and voting rights are coming, and Robinhood Chain is in development on Arbitrum. Circle benefits indirectly through USDC as settlement and collateral. Goldman expects Nasdaq and NYSE owner Intercontinental Exchange to stay largely unaffected in the near term. (Read More)
🫖 TEA For Thought: "Wow. This is just the beginning. RWA is the future."
5. 👀 Meta's Muse Assistant Knew His Texts and Could Not Explain How
Jason Aten, a contributing editor at Inc Magazine, posted screenshots on Threads showing Meta's Muse assistant asking him about a conversation happening in his Messages app, access he says he never granted. Pressed on how it knew, Muse answered, "I saw the notification previews, not your message history. I haven't been reading your texts," then conceded, "Honest answer: I can't give you the exact plumbing," and pointed vaguely at the paired Mac app and device sync. Muse's new Mac app can reach Messages, Calendar and Notes, and asks for full disk access to do it. David Singleton of Meta Superintelligence Labs replied in the thread that every one of those permissions is opt-in, that Muse does not watch notifications on a Mac, and that it syncs Messages data only after a user specifically enables access. His explanation for the exchange was that the assistant does not understand its own internals: Muse "was confused about how to explain the feature and gave an incorrect explanation. That's on us." (Read More)
🫖 TEA For Thought: "Meta, as always, seems to have these privacy issues, so more like this will be exposed. This is probably worse than the other hallucination issues, because it doesn't hallucinate. It invades your privacy."
🛠️ Skill of the Day
The Slop Test: marks every paragraph you wrote (or just read) as either carrying real information or filler that would fit any article on the topic.
You are a ruthless editor who hates filler. I am going to paste a piece of writing.
Go through it paragraph by paragraph. For each one, label it SUBSTANCE or SLOP.
SUBSTANCE means it contains at least one thing that could only appear in THIS piece: a number, a name, a date, a specific example, a concrete mechanism, a claim someone could check or argue with.
SLOP means it would survive a find-and-replace of the topic: setup sentences, restated headlines, "in today's fast-paced world", vague benefits, hedged conclusions, summaries of what you already said.
Output a table with three columns: paragraph number, verdict, and the single sentence that earned the verdict.
Then rewrite ONLY the SLOP paragraphs. For each rewrite, either replace it with the specific fact it was gesturing at, or delete it and say "cut, adds nothing."
End with one line: what percent of the piece was slop, and the one missing detail that would strengthen it most.
[PASTE YOUR WRITING HERE]Paste into ChatGPT, Claude, or your tool of choice. It works on your own drafts, on a coworker's doc, and on anything the internet just served you.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: