Three pieces today don’t cite each other and argue the same thing. In Fast Company, Amit Joshi reports that American Anthropic customers are swapping the frontier model for cheaper options, and that Uber started sending simple tasks to a cheap model. Tyler Cowen, from the theory side, describes intelligence and tacit knowledge as near complements: when intelligence becomes abundant, the value of what stays scarce goes up. And L.L. Bean spends $50m on a store with a trout aquarium and almost no screens. Putting the three together is my reading: when the input you sell gets cheaper, the price moves to what takes years to build — the customer relationship, the place, the silhouette of a fries box that McDonald’s bets is recognizable from its shadow alone. The headline on the 7th said that whoever has no channel has no yardstick either; today’s adds that they have no price either.
What tightens this math is the price of money. The long-term US rate sits just below its highest since 2007, and Axios puts the capital demand of the AI buildout itself among the forces pushing it. The frontier race only works if the most expensive model holds a margin above “good enough”, and with high rates that margin has to show up sooner. On the same Saturday, Altman said OpenAI won’t go public in 2026 and gave safety as the reason. From what was published, there’s no way to separate the weight of the justification from the weight of the window. I’ll only note that a private company doesn’t have to show margin every quarter, and that the premium premise started being questioned in public on the same weekend.
product and growth
L.L. Bean bets on the physical store — L.L. Bean put $50m into renovating its Freeport, Maine flagship: 160,000 square feet, aquariums with native trout, a fly-fishing demo at the entrance and almost no screens. The store was already the state’s second-biggest tourist attraction, with more than 4 million visits a year, and the renovation leans into that: what gets sold first is the visit, and the company put someone who built a career in stores in charge (Greg Elder went from VP of stores to chief retail officer before becoming CEO). The detail that talks to the rest of the day is the origin: the brand started in 1911 mailing a flyer to the list of Maine hunting license holders and spent 80 years with a single store, meaning the customer relationship was always the asset, and the store is the new channel for that relationship. The piece shows no return on the investment; for now it’s a bet, not a case.
brand and ip
McDonald’s uses only the shadow of its products — TBWA\Neboko put four pieces on Dutch DOOH in which fries, Happy Meal, McFlurry and McNuggets appear only as a shadow cast in the sun, on the beach, on the street, on a sports field, with no color and no headline. It’s the distinctive assets thesis turned into media: the shape of the packaging would carry the brand on its own. The detail I wouldn’t let slide is that the logo is still in the bottom right corner, so the piece doesn’t prove the shadow is enough; it shows the brand trusts the shadow, but not without a net. No measurement has been published.
market and capital
Whoever owns the customer beats whoever has the best model — Amit Joshi starts from an observation with no number attached: many American Anthropic customers are choosing cheaper models over Fable 5, on the eve of what could be the biggest IPO in history and right after the company’s first adjusted operating profit, and the article doesn’t say how many. The concrete case is Uber, which burned through a year’s worth of tokens in four months, started sending simple tasks to a cheap model and saving the expensive one for the hard stuff, and saw AI usage grow more than ninefold with no matching rise in spend. Joshi’s thesis is that customers will buy outcomes and build model portfolios, and then the frontier model stops being the product and becomes an input; his recommendation is that the labs become the platform that picks the model per task. My question is the one he doesn’t face: why would the customer hand routing over to precisely whoever sells the most expensive model.
OpenAI delays its IPO and cites safety — Altman told Fortune that OpenAI won’t go public in 2026 and that it would rather do its safety and alignment work as a private company. The market was counting on IPOs from both this year, and Axios notes the decision leaves the ball with Dario Amodei, who on the same Saturday called for AI development to slow down. Separating things: safety is the stated reason; the window, with the long rate where it is and the frontier premium in doubt, isn’t a great one either. How much each weighs, the piece doesn’t let you say.
The bill for American debt comes due — The 10-year Treasury hit 4.97% on Friday, one percentage point above the end of February and just below its highest since 2007; the 30-year fixed mortgage went to 7.08%. The immediate trigger was August inflation, with gasoline at $4.29 a gallon accounting for more than a third of the rise, and a Fed rate hike is expected this week. But the long rate is set in the global market and by bigger forces: the government spends about $2tn a year more than it collects, debt is close to 100% of GDP, and Neil Irwin puts the capital demand of the AI buildout among what’s pushing long rates up worldwide. That’s the mechanism tying this item to the two above: the frontier race competes for the same money it helps make more expensive.
reading
Backtest of Piotroski’s F-score on the S&P 500 — A teaser, via Quantocracy’s summary, for a Quanter Lab test: the F-score (nine yes-or-no questions about the last two annual balance sheets, one point per yes, buy whoever scores 8 or 9) applied to the S&P 500 with point-in-time data from 2000 to 2025. The excerpt only gets as far as the origin: Piotroski designed the score in 2000 for the cheapest fifth of the market by book-to-market, where a positive return on assets was information, and there the high scorers returned 7.5 percentage points a year above the group; the S&P 500 result doesn’t appear in what I read. The idea that sticks, with or without the number, is about method: a signal calibrated on the cheap tail of the market has no reason to hold among the largest companies, and point-in-time means testing only with the balance sheet and index composition that existed on each date. For people building product, it’s the trap of the metric imported from another segment.
who wrote
Tyler Cowen and a simple model of AI-aided growth — Cowen swaps Solow for an economy with two factors: intelligence (the formal kind, the kind you measure in an eval) and Polanyian knowledge, the tacit know-how of place, custom and habit, which no AI yet learns just by walking into the office. Since the two are near complements, the current intelligence shock should raise returns, employment and real wages in the sector that stays scarce, only slowly, because that sector is messy by nature. The consequence he highlights is that whoever controls the intelligence sector will have less power than it seems, because the complements are missing. For people building product, it’s the macro version of Joshi’s Uber case: intelligence becomes an input, and what the customer pays for is knowing how that particular office works.
stalled sources
First Round Review (321 days), Gurwinder (259), Calculated Risk (245), Elad Gil (146), Y Combinator Blog (89), Anti-Mimetic (60), Acquired (35), Adjacent Possible (35), Kyla Scanlon (31), Sherwood News (27), Collab Fund (27), Commoncog (20), Granted (16).