The Adjacent Brief

TL;DR: South Korea is committing roughly $1 trillion in state and industrial capital to domestic AI infrastructure, according to a Semianalysis breakdown of the program. Pennsylvania Governor Josh Shapiro, until recently one of the loudest data center boosters among US governors, called some projects "predatory" and imposed new guardrails. Instagram tightened its treatment of undisclosed AI profiles, relabeling them and cutting their algorithmic reach.

Worth Reading

• Trump turns data center siting into a midterm loyalty test (paywall) — Communities that reject data centers will end up "backwards and poor," per the President; the local permit fight now has a national partisan frame.
• The endpoint becomes the enforcement point for corporate data — Security is moving below the OS as AI PCs take on data protection, which means the laptop refresh cycle is now a security purchase.
• Instagram stops calling synthetic accounts "creators" — The label change from "AI creator" to "AI-generated profile" is small wording with large positioning consequences.
• Drivers are rejecting car features they have to learn — J.D. Power finds owners want technology they can ignore, a useful corrective for anyone shipping an in-car AI assistant.
• Mandates don't produce AI transformations, and the org chart is why — Julie Zhuo on why top-down adoption targets generate compliance theater instead of changed workflows.
• a16z argues platform decay is a story people tell, not a mechanism — Worth reading even if you disagree; the counterargument sharpens what "enshittification" actually claims.
• Perimenopause became a venture category with creator-led distribution — WIRED on hormone therapy startups using influencer channels and medical framing to reach millennials.

Brand & Growth

Korea is buying compute the way it once bought shipyards

The program trillion-dollar sovereign AI program Semianalysis details is industrial policy aimed at owning the stack Korea currently rents. Korean firms already supply the high-bandwidth memory that every frontier training run depends on, so the country is building outward from a chokepoint it controls rather than starting from zero. This sits alongside a pattern on the enterprise side: workloads migrating from public cloud back on-prem for cost, control, and sovereignty reasons. For anyone with a multinational vendor list, "which model" is becoming a question with a geography column attached, and procurement teams in Seoul, Brussels, and Riyadh will start asking it before the CMO does.

When the platform makes the decisions, measurement is the job

Google Ads keeps absorbing decisions that used to be a media buyer's craft, and Search Engine Journal's argument is that the surviving skill is proving what the automation actually caused — incrementality testing, geo holdouts, clean first-party attribution. That reframes the adjacent piece on restructuring marketing teams and budgets for AI search, which makes the case that visibility in AI-generated answers is a reallocation problem rather than a headcount problem. Both point at the same budget line: money moving out of execution roles and into measurement, research, and content that models will cite. The uncomfortable implication for agency relationships: if the platform sets bids and the model writes the copy, the retainer has to be justified by evidence, and most retainers currently aren't.

Connected World

The bottleneck for AI buildout is consent — silicon is no longer the limiting factor.

Shapiro spent two years courting hyperscalers and now calls some data center projects "predatory" while imposing guardrails — and the most instructive detail in the Times piece is that western Pennsylvania, the part of the state with the least to lose, still wants the projects. The fight is affluent counties with rising utility bills against post-industrial ones trading ratepayer exposure for a tax base. That explains the otherwise odd math of an AI company paying $75 million to put its name on Texas Tech athletics. Regional goodwill purchased in the same states where these firms need land, water, transmission interconnects, and a friendly zoning board functions as pre-emptive permit strategy, making the cost per acquisition metric beside the point.

The hard part of physical AI is a supply chain, not a model

Rare-earth permanent magnets sit inside every actuator in every humanoid robot, and TNW's argument is that the magnet, not the model, is the binding constraint on robotics scaling, with China controlling processing and Europe's Critical Raw Materials Act still years from producing meaningful volume. Anyone forecasting robot deployment curves off model capability is forecasting the wrong variable. Whether Meta can scale the humanoids already doing cable-plugging and server resets in its live data centers depends on magnet output in Ganzhou, and no benchmark will answer that question.

The New Consumer

A third of UK adults just priced the streaming bundle themselves

Roughly 33% of UK adults used illegal streaming services in the past three months (paywall), costing sports, TV, and film an estimated £1.4B a year, per the FT. That's behavioral data, not a survey about intentions, and it measures something no willingness-to-pay study captures: the point at which fragmentation stops being an annoyance and becomes a reason to leave. Piracy at that scale is a pricing signal from the segment that has already subscribed to three services and hit a wall on the fourth, most acutely in live sports, where exclusivity is the entire product strategy. The industry read will be enforcement. The commercial read is that re-bundling is worth more than litigation.

Familiarity with AI is producing skepticism rather than comfort

TU Darmstadt's AI Monitor found that 43% of Germans who rate their AI knowledge as very strong report negative expectations and renaming the account type. The label is what gets covered; the distribution penalty is what changes behavior. Meta has no interest in banning synthetic accounts, since they generate content supply, but it has a strong interest in advertisers not discovering that their influencer buy went to a character. For brands running creator programs, verification of a human on the other end has become a media-buying checklist item.

Culture & Signal

Ranking systems, rather than rules, settle the slop fight

Derek Thompson makes the cultural case that the internet is drowning in machine-written text and that this is worth resisting — a normative argument about what a public information commons is for. Ben Thompson, working the institutional side, reads Meta's content settlement as evidence that our regulatory frameworks are badly matched to the governance problem they're being asked to solve, and proposes thinking in terms of distribution rather than speech. The mechanism is clear: courts settle money, legislatures settle jurisdiction, and the actual volume of AI content people encounter is set by whoever tunes the feed. Instagram's throttling decision, noted above, is a more consequential act of content policy than any settlement signed this year, and nobody voted on it.

Commerce Rewired

Impressions are inventory; advertisers are the business

The case Lewis Lin poses — local ad impressions up 35% while local business sign-ups fall 8% — is a good diagnostic because the obvious answer is wrong. Rising impressions with falling advertiser counts usually means supply inflation (more feed surface, more AI-generated content to sell against) meeting a shrinking or consolidating buyer base. Impressions grow, price per impression softens, and revenue holds for a quarter or two while the customer base thins underneath. Net new advertiser count and second-month retention are the true health metrics for the platform; delivered volume is a lagging indicator of inventory rather than business strength. A platform reporting impression growth without advertiser growth is describing how much space it has to sell while leaving unanswered how many buyers it has actually won. actual business performance unexamined.

Machines & Minds

An old security category is getting repriced for agents

Data loss prevention was a compliance checkbox that nobody loved — rule-based pattern matching, endless false positives — and it's getting a second act as vendors swap regex for contextual understanding, with CrowdStrike folding Jazz into Falcon. Follow the buying logic rather than the product news: enterprises deploying agents that touch source code, claims data, and cloud infrastructure need to know what left the building and why. Forrester pushes the same idea further, arguing that intent — not prevention or detection — is where agentic security has to operate, because an agent modifying cloud resources produces legitimate-looking actions faster than monitoring can evaluate them. The budget consequence is concrete: security spend follows agent deployment on roughly the same curve, and CISOs who approved agent pilots in Q1 are writing the tooling checks now.

Persuasion is showing up on benchmarks before it shows up in harm data

Recent evaluations show models improving measurably at persuasion, deception, and social engineering, and Transformer's read is the disciplined one: the capability is real, the real-world harm is still unquantified. Benchmark persuasion under lab conditions with a captive participant is not the same as changing behavior at scale against distracted people with priors. Worth holding both: nobody has demonstrated the mass-manipulation scenario, and the capability curve is pointed in an uncomfortable direction. If intent detection is where agentic security is heading, the harder version of the problem is that the target of the manipulation is the human approving the agent's request.

15 articles across 6 themes · 13 sources · Powered by Folo + Claude