An internal Microsoft document, redacted until this week in the New York Times case, says in one line what the industry burns entire press releases trying not to say: “LLMs are a product that destroys its supply chain.” In the same batch, another internal text describes the company’s content strategy as a “doom loop” that will hurt both model performance and the web at once. This isn’t the author filing the accusation: it’s the buyer of the raw material describing the mechanism in writing, years before the bill arrived.
The curious part is that the same kind of mechanism shows up today, measured somewhere else, with no relation to AI at all. The paper by Kurth, Eisler, Rej and Bouchaud on the death of short-term trend-following argues that trend was never a static anomaly somebody harvests: it was a feedback loop where the trade pushes the price and the price sustains the signal for the next trader. Break the loop and profit and signal die together. The difference between the two cases is that Bouchaud names the variable and measures it; Microsoft wrote a memo. The rest of the day is about who decides what to do with that dependency: the creator starting to sell “I got cited” instead of audience, Oreo finding out its brand became raw material for somebody else’s campaign, and, on the other side, three cases of people who still control their own chain and are putting a price on it.
launches
Meta One — Meta packed everything into a bundle with a short name and six tiers, from $2.99 to $499 a month, and already reports 15 million subscriptions and trials. The real decision isn’t the bundle, it’s what’s buried in the business plans: a link in an organic Instagram post becomes a paid feature, four a month at $49.99 and twelve at $499. In other words, the price isn’t on distribution, it’s on the exit — the platform created the friction and now sells its removal in countable units. The other detail that changes work: a WhatsApp Business MCP server lets Claude, Cursor or ChatGPT create the account, verify the number and build the templates, which wipes out several days of setup some agencies were billing as a service.
product and growth
Lucras, 99Pay’s mascot — the Acelerador de Lucros campaign, out of Crispin, got a mascot named by public vote, with digital media, influencers and broadcast TV support in the Northeast and the South. The mascot exists to carry a single number: a yield above 130% of the CDI, against the savings account, with any amount earning even on weekends and holidays. The 88% approval cited comes from research 99Pay itself commissioned from On The Go, which is promotion, not evidence. My builder’s caveat: a rate isn’t positioning, it’s an offer — when the CDI turns, Lucras is still there and the argument he carries is gone, and that’s when you find out whether the brand asset was the mascot or the rate table.
brand and ip
the unredacted documents in NYT v. OpenAI/Microsoft — the publishers’ motion led by the Times surfaced what executives at both companies were writing internally about training on someone else’s archive, including a Microsoft applied science director predicting that millions of people would call it “an astonishing theft of unprecedented proportions”. Microsoft replied that this is one employee’s view and not the company’s, which is the only possible answer, but it doesn’t erase the rest: internal communications identifying the risk that generative AI puts exactly the people who produced the data out of work, Nick Turley acknowledging an existential threat to publishers and “largely substitutive” products, and OpenAI employees getting around paywalls to scrape sites. For anyone sitting on the other side of a licensing table, this isn’t gossip: it’s the negotiating floor, because the counterparty has already documented that it knows what it’s buying and what it’s worth. The part that matters to product is the one in the policy document: a final product that threatens the economic base of its essential suppliers is a structure that funds itself today by consuming tomorrow’s raw material.
Oreogate — a kindergarten parents’ group chat about banning Oreo from the lunchbox turned into an AI-generated song and then into dozens of videos by a creator with 237,000 Instagram followers, plus third-party t-shirts and mugs. Mondelez told the Wall Street Journal it’s “thrilled” and that it wasn’t behind the phenomenon, and licensing lawyers interviewed by Licensing International recommend a risk assessment instead of a reflex, noting that the enforcement itself becomes the story — the precedent cited is Warner’s cease-and-desist that made Stew Leonard’s rewrite its Ted Lasso biscuit sign in 2022. The new element, and one no playbook covers, is that Suno confirmed the creator was a paid partner producing content with its music generator: Oreo’s brand showed up as raw material in a paid campaign for another product, with no contract, no brief and no fee. Monitor and wait is still the right call for this case, but the structural question stands: when the creative input is your brand, the one who pockets the awareness is the one who has the distributor.
media and attention
creators learning to sell AEO — the pitch is shifting from “my pasta video went viral and got this many views” to “my video about PFAS in water shows up when you search for it”, according to Crystal Duncan of Tinuiti, who says the mention now comes up unprompted in media kits. It’s the supply side of what this daily covered yesterday from the advertiser side: if the model’s answer is the new inventory, whoever feeds the answer wants to get paid for it. The problem is that “I got cited” is still an anecdote about one search, not a metric — and James Chandler, of IAB UK, gives the right warning: showing up in an AI answer isn’t the same as having caused the recommendation, and without that distinction “AI influence” becomes a vanity metric. The lesson for anyone building an offer: the metric the buyer doesn’t yet know to ask for is exactly the one worth building first, but only if it survives a causality question.
market and capital
“Is Trend Still Your Friend?” — the July 2026 paper by Kurth, Eisler, Rej and Bouchaud, reviewed by Larry Swedroe, starts from four facts: the break in short-term trend in 2008-09 was abrupt, speed-dependent, uneven across asset classes and never recovered, even with better liquidity and fewer CTAs in the market since 2018. They rule out capacity (industry AUM only peaked in 2022, far too late to explain it), electronification (gradual, while the break was sudden) and a change in the interaction with order flow, and land on a variable that separates cleanly: the contract’s tick size normalized by volatility. In small-tick contracts, fast trend P&L went to zero after 2008; in large-tick ones, it kept earning at close to the old rate, and sorting by liquidity or by asset class doesn’t reproduce that split. The explanation is the broken loop: with market making taken over by HFT running a zero-inventory mandate, whoever used to absorb the predictable directional flow from CTA rebalancing disappeared, and in a sparse book there was no depth left to execute against. This is not an allocation tip of any kind; it’s a rare case of an anomaly whose existence depended on one specific participant staying at the table.
Revolut’s dual listing — Nikolay Storonsky told Les Echos that the fintech, valued at $115bn in a secondary sale, is studying a listing in New York and London, after saying in 2024 that the English exchange “can’t compete”. The stated preference is still the United States, on the simplest argument there is in book building: a big market with many buyers fighting over the shares versus a small market with few. The context that explains the reopened door is operational, not sentimental — a preliminary OCC banking license, a Stamford headquarters with $95m and around 160 people to launch in 2027, plus licenses in France and the UK, and British ministers lobbying. The IPO still has no timeline before 2028, which is plenty of time for the stance to change again.
reading
the four steps before launching something new — Eric Athas, deputy editor on the New York Times newsroom development team, wrote a book called Saying No to New and applies the thesis in reverse: whoever gets good at turning down the wrong new thing gets good at spotting the right one. The step with operational detail is the second: every project on the team starts with a six-field proposal — the problem in one sentence, what you’re going to do and why, the expected outcome in one sentence, who runs it and who needs to be informed, when it starts and when it ends, and the step by step. Then comes a pilot with two desks before any rollout, which is the same remedy against planning fallacy used in product and almost never used in internal training. What this changes for anyone pushing an AI tool into a newsroom: a colleague’s skepticism isn’t resistance, it’s a request for justification, and anyone who can’t fill in the six fields didn’t have a project.
who wrote
Collab Fund’s thesis on Dr. Max’s — in a market where nearly every thesis is a bet on AI, the fund wrote up why it’s backing a family doctor from Mendocino who spent five years prototyping a product in his kitchen with ingredients from his own farm. The diligence argument is the origin of the demand: more than 30,000 patient consultations asking for the same thing before a product existed, a formula tuned by patient feedback, and less than $10k of marketing before it became the best-selling brand in children’s body care at Erewhon. Worth marking the denominator: Erewhon is a small, extremely expensive retailer, so that’s a cohort signal, not a market one, and the comparables cited (Burt’s Bees sold for $925m in 2007, Coterie for more than $1bn, Eucerin and Aquaphor with €1.5bn in annual revenue) measure the ceiling, not the probability. What it teaches, on the day Microsoft wrote that the LLM destroys its supply chain, is the contrast in structure: here the same person is the origin of the demand, the supplier of the raw material and the owner of the initial distribution.
the rest in one line
- Still in the Digiday briefing: Judge Brinkema rejected the breakup of Google’s ad tech in a 106-page opinion, but requires the company to stop tying its ad server to AdX real-time bids and to accept a compliance monitor for six years, instead of the 15 the DOJ asked for; Google will appeal the liability finding.
- From the same briefing, via Tube Filter: TikTok says 94% of luxury resale revenue on TikTok Shop in the US already comes from live streams, where you can show the item’s condition and answer questions before someone spends big.
- And the number to keep: £1.2bn is the forecast for advertiser investment in creator partnerships in the UK this year, the first time at that level.
stalled sources
First Round Review (326 days), Gurwinder (264), Calculated Risk (250), Matthew Ball (214), Elad Gil (151), Y Combinator (94), Anti-Mimetic (65), Adjacent Possible (40), Kyla Scanlon (36), Sherwood News (32), Granted (21), Snaxshot (16), Benedict Evans (15), Investment Idiocy (15).