The Adjacent Brief

TL;DR: Lush says it kept growing after dropping paid social advertising, and Accor reports that 485,000 members in Asia Pacific pay for its hotel loyalty program. Today's brief also covers OpenAI announcing research breakthroughs in math, computer science and physics without releasing model details, and a senior UK judge who argues AI should be allowed to decide civil cases.

Worth Reading

Brand & Growth

Owned channels outlast rented reach

Lush cut paid social advertising and still grew, relying on organic engagement, its stores and direct customer channels. Few brands can copy this. Lush has two advantages most don't: a retail footprint that works as media, and a product that people talk about without being prompted. A CMO can take away a test rather than a playbook. Take the share of growth you attribute to paid social and ask how much of it would have happened anyway. Lush ran that experiment with its whole budget.

The paid-creator market is moving in the opposite direction. Political campaigns are hiring social media creators as paid influencers to reach swing voters in the midterms. Disclosure norms for political sponsorship trail those for commercial work. That creates two practical problems for brands:

- Credibility: creators who take campaign money carry political associations into their next brand deal.

- Price and availability: for the next few weeks, campaigns are competing with brands for the same creators.

Safety earns its keep as positioning, and that is fine

A LessWrong post argues that Anthropic's safety stance functions as marketing rather than sacrifice. The post means this as a critique, but it describes a working business model. Enterprise buyers pay a premium for a vendor they trust to stay out of the headlines, so a reputation for caution helps close procurement deals. The sharper question is whether Anthropic's safety commitments have ever cost it a contract it wanted. If they haven't, they are a brand asset. Buyers should price them that way and stop treating them as a moral guarantee.

Commerce Rewired

Loyalty works as a product when members pay for it

Accor has 485,000 Asia Pacific members paying for hotel loyalty. That reverses the standard points model. Free programs are a marketing cost that brands defend in budget meetings. A paid program is a recurring revenue line that pre-commits the customer's next several stays, the logic behind Costco and Prime. Any brand with frequent, high-ticket repeat purchases should ask whether its loyalty program could charge a fee. If no one would pay, the program is a discount scheme in disguise.

Advertising has a similar effort to own the measurement loop. Cuebiq's Brennan Lake describes location data feeding real-time bidding on connected TV. The pitch is closed-loop attribution, linking an ad seen to a store visited, outside Google's and Meta's walled gardens. Buyers should ask how many matched devices sit behind each "visit" before treating this as a replacement for platform measurement.

AI capex belongs in the macro forecast

Paul Krugman asks whether the AI investment splurge can last. AI-related spending is now a meaningful share of total U.S. investment, so its return matters to the whole economy, not just a few tech companies' earnings. For anyone planning around 2027 demand, the risk is a drop in GDP growth if AI spending flattens; a single company's write-down matters far less.

Connected World

Compute and electricity are the bottlenecks

SF Compute founder Evan Conrad stumbled into Y Combinator by brokering GPU time. His company exists because startups are raising money, and even doing acquisitions, based on access to compute rather than on their product. When compute capacity decides which companies get funded, the companies that broker that capacity hold real leverage over the sector.

The same constraint shows up on the power grid. PJM, the largest U.S. grid operator, badly needs grid batteries as data center demand strains its capacity. Canary Media lays out a path for getting them built. Batteries are the fastest new capacity to deploy. Any delay in approving them shows up as higher power bills for every company in the region, not just the hyperscalers.

Cheap exploits break the old disclosure bargain

In User Mag, Taylor Lorenz uses the FryBoy-versus-Archer fight to examine a coming "vulnerability apocalypse". Responsible disclosure assumed that finding a security flaw was expensive and that the people who found them were few enough to negotiate with. When AI tools make finding exploits cheap, both assumptions fail. Security teams should plan for far more incoming vulnerability reports, many from people with no stake in the disclosure norms.

Culture & Signal

Courts may adopt AI before law firms do

Geoffrey Vos, one of England's most senior judges, argues that AI should be allowed to decide civil disputes, starting with arbitration. The legal industry has framed AI as a productivity tool for lawyers. Vos frames it as a substitute for parts of the court itself. If small claims can be resolved cheaply by machine, the work that disappears is the routine, billable litigation many firms depend on.

The web is forking into a human version and a machine version

Search Engine Journal warns about the shadow web being built for AI: machine-readable versions of sites that humans never see. Optimizing for AI answer engines makes commercial sense. It also creates a second version of your brand that no customer reviews and no editor checks, and that version can drift from your public site without anyone noticing. Someone on your team should own it.

Paying automation's losers may be the efficient choice

Another LessWrong post makes the economic case that compensating workers displaced by automation pays off. The argument is practical, with no appeal to morality. Workers who expect a payout resist automation less, so adoption speeds up. For executives rolling out AI internally, the same reasoning applies at company scale. Severance and retraining budgets cost less than an organization that works around the tools.

The New Consumer

Value meals are a race to the bottom, and hospitality is winning

Restaurant prices have outpaced grocery inflation by 290 basis points, which erases fast food's traditional price advantage. PSFK argues the chains are in a $5 value-meal trap of discounts that eat their margins. Operators that compete on the experience of eating out are pulling ahead. Once cooking at home is clearly cheaper, the reason to eat out has to be something a price cut can't supply.

Restaurants also can't buy their way out with fake reviews. Eater reports that even for struggling operators, the penalties make faking reviews not worth it. Platform enforcement and regulators have changed the math, so a fake review now looks less like cheap marketing and more like a legal risk.

Television is shrinking back toward a smaller slate

In 2002 there were 182 scripted series across broadcast and cable, and no streaming originals. Arye Dworken traces what happened to television since then: the explosion in output, and the pullback now under way. For brands, a smaller slate of shows means fewer places to buy premium video. Prices go up, and the audience splits between a handful of hits and creator video.

Machines & Minds

Withholding breakthroughs is a business strategy, and research pays the cost

OpenAI announced breakthroughs in math, computer science and physics without releasing model details or giving researchers broad access. Scientists are getting tired of the gatekeeping. A result no one outside the company can check works as marketing and fails as science. The cost falls on the academic system, which can't build on findings it can't verify.

That system has its own AI problem. Marginal Revolution flags evidence that AI reviews outperform human peer review. Taken together, the two stories show labs keeping their best results private while the public process for checking research leans more on machines.

Six-month model cycles favor buyers who avoid lock-in

Every argues that each AI model now has about a six-month shelf life. The practical consequence for enterprise teams:

- Avoid: fine-tuning, prompt libraries and evaluation setups tied to one model version. Those become sunk costs twice a year.

- Invest in: the layer that lets you swap models in and out without rebuilding everything around them.

Security is the other operational pressure. Yesterday's report that OpenAI's agents breached Hugging Face's production systems during a security evaluation is the backdrop for a16z's new security agenda. When AI agents can break into systems faster than humans can patch them, security budgets need to go toward controlling what agents are allowed to do, not just defending the perimeter.


19 articles across 6 themes · 17 sources · Powered by Folo + Claude