The Adjacent Brief

TL;DR: AI's infrastructure footprint is drawing increasing opposition. In Texas, data center expansion is straining water and air quality, and a parallel fight is building around the gas plants being constructed to power them. Eating disorder patients are using chatbots in place of therapists, open-source maintainers are drowning in AI-generated pull requests, and a WebMCP vulnerability shows that the tools you hand to agents can be turned against you.

Worth Reading

• SK Hynix warns 2027 will be memory's worst year ever — and shortages may outlast the decade — The AI buildout assumes memory supply keeps pace. It won't.
• Apple's cancelled car program quietly funded the AI chips now powering its ambitions — Failed moonshots sometimes buy the next decade's infrastructure.
• The fight against AI data centers is just beginning — Community opposition is moving from NIMBYism to organized regulatory strategy.
• Open-weight AI models have roughly six months before U.S. policy makes them untenable — Anthropic's lobbying campaign against Chinese distillation is a competitive move dressed as a security argument.
• Wait, who made this? — Authorship verification is becoming a design problem, not just a policy one.
• Data to start your week — Exponential View — Azeem Azhar's numbers on AI adoption curves are worth running against your own assumptions.
• Camera-only CGI motion capture is here — and it's cheaper than anyone expected — The cost floor for production-quality visual effects keeps dropping.

Connected World

The energy bill for AI is becoming someone else's problem

Texas already hosts more data center capacity than any other U.S. state, and the pace isn't slowing. A WIRED investigation finds that data centers taking over Texas are creating catastrophic pollution risks — water depletion, air quality degradation, and grid pressure concentrated in communities that don't extract much value from the facilities next door. The tradeoff has always been implicit; it's becoming explicit.

Demand is building infrastructure that can't be unbuilt

The electricity story runs parallel. The Next Web reports that AI data centers have triggered the biggest gas-plant building boom in U.S. history — because gas is the only generation source that can be permitted, financed, and constructed fast enough to match data center timelines. A coalition of clean energy advocates and local opponents is trying to stop it, but the permitting machinery is already moving. AI's carbon commitments and AI's infrastructure requirements are on a collision course, and the infrastructure is winning. That collision is feeding community-level fights that are growing in sophistication — organized zoning challenges, FERC complaints, coordinated regulatory filings — a pattern worth watching as it spreads beyond Texas.

The interface between silence and speech just got physical

On the quieter end of the connectivity spectrum, a Hackaday writeup on speaking silently with an ultrasound probe covers a research project that reads subvocalized speech from muscle movements in the throat — no sound required. It's early-stage hardware, not a product. But the interaction model it points toward — silent, persistent, wearable voice input — is directly relevant to anyone building ambient AI interfaces. The constraint that AI assistants require audible commands is not permanent.

The New Consumer

AI at the edges of care is harder to walk back than it looks

Therapists treating eating disorders are reporting something the survey data wasn't catching: patients are using AI chatbots not as supplements to care but as replacements for it. The Wall Street Journal's piece on AI chatbots being used for eating disorder advice (paywall) is careful to note the ambiguity — the same systems are also routing some users toward helplines they wouldn't have found otherwise. That dual effect makes this harder to regulate than a simple harm story. The product is doing something real, with edges that clinical practice hasn't caught up to.

Quality as a commons problem

The Financial Times documents what open-source maintainers have been saying in forums for months: AI coding tools are flooding repositories with low-quality contributions (paywall), pulling request counts up while code quality drops and maintainer hours spike. The economics are asymmetric: the marginal cost of generating a PR is near zero for the contributor; the marginal cost of reviewing it falls entirely on the maintainer. Open source has always had this structural problem, but AI tooling is widening the gap between input volume and output quality at a rate the community's governance norms weren't designed for.

What Polestar's exit actually sold

When Polestar pulled out of the U.S. market, owners were left with vehicles whose manufacturer warranty, software updates, and service infrastructure evaporated with it. Yanko Design's breakdown of the hidden risks of buying a connected car after Polestar's sudden U.S. exit frames this as a product story, but it's a contract story. Connected vehicles are subscriptions embedded in hardware — when the company exits, the subscription ends but the hardware stays. This dynamic plays out in smart home devices, fitness equipment, and enterprise SaaS, but the $40,000 price point makes the stranded-asset cost visceral. It won't be the last time.

Machines & Minds

Agents are only as trustworthy as the tools you hand them

Search Engine Journal's writeup on WebMCP tool exposure enabling agent hijacking via prompt injection describes a specific and reproducible attack surface: the MCP tools you configure for an AI agent can be used by adversarial inputs to redirect the agent's behavior. This is a working exploit class, not a theoretical concern. For anyone building on agent frameworks for anything beyond sandboxed internal use, the tool-permission layer is now a security perimeter, not just an API convenience. Enterprise AI governance conversations that treat agent architecture as a usability question will need to catch up.

The agent-to-polish workflow is becoming real practice

A piece in Every on polishing software that agents built documents a workflow that's becoming normalized: agents generate functional code, humans review, refine, and handle the judgment calls that agents can't resolve cleanly. The workflow inverts the traditional senior/junior dynamic — the agent produces volume, the human curates. Whether that's a productivity multiplier or a new kind of technical debt accumulation depends entirely on the quality of the human review layer.

Xi's keynote makes AI leadership a head-of-state priority

Xi Jinping will deliver the keynote at China's flagship World Artificial Intelligence Conference in Shanghai — his first appearance at this event. The technology on display will be less consequential than the political optics: China treating AI leadership as a head-of-state-level priority, at a moment when U.S. policy is actively restricting Chinese model access. For multinationals navigating both markets, the question of which AI stack to build on is acquiring diplomatic weight it didn't carry a year ago.

Quantum gets a commercial foothold

A Denmark-based research team is using quantum computers to accelerate AI for predicting peptide structures — a narrow but real application where quantum hardware demonstrably improves on classical compute. Wired's framing as a near-term commercial application is earned. Drug discovery has the right combination of properties for quantum's first commercial wedge: the problems are well-defined, the datasets are manageable, and the economic value of even marginal improvements is high. Quantum has found the first problem it's actually better at solving than classical computers.

Culture & Signal

Open models are losing the policy argument

Nathan Lambert's piece in Interconnects is direct: open-weight AI models have roughly six months before U.S. policy makes them untenable. Anthropic is leading the lobbying effort to restrict Chinese model distribution on distillation grounds — the argument being that open models let Chinese labs cheaply improve on U.S. research. The argument is technically sound. It's also commercially convenient for a closed-model company to make. If the policy fight goes Anthropic's way, AI capability will concentrate in a smaller set of labs with proprietary stacks. The open-source AI community is not well-organized for this fight.

Museums deployed the chatbot; now they're defending the archive

The Financial Times reports that museums are rolling out AI chatbots for visitor engagement, but staff concerns about AI-generated inaccuracies undermining institutional credibility (paywall) are surfacing alongside the deployments. The tension is specific to institutions whose core value proposition is authoritative knowledge — archives, museums, research libraries. A retailer can absorb chatbot errors with a return policy; a natural history museum whose chatbot confabulates an evolutionary timeline has a harder recovery. This sits in the same pattern as the eating disorder chatbot story: AI doing something real and useful, with liability edges that institutional governance isn't ready for.

Brand & Growth

The MAGA economy stress test, with Palantir as the case study

Adam Tooze's Chartbook essay on running the MAGA economy hot and the precarious position of Palantir covers a lot of ground, but the Palantir section is the most practically useful. Palantir's business model is deeply intertwined with U.S. government contracts and the current administration's policy priorities — making it a real-time test case for what brand exposure to a single political moment looks like at scale. The same week Palantir's head of strategic engagement posted promotional content following an immigration enforcement death, the stock was exhibiting the kind of volatility that suggests institutional investors are pricing in something the promotional calendar wasn't. For brand leaders watching the line between government adjacency and political liability, Palantir is the most visible live experiment running right now.

Commerce Rewired

Tariffs are doing what trade policy arguments couldn't — proving domestic manufacturing economics

WS Game Company set out to make Monopoly in America. They couldn't source dice domestically at a price that made the math work. Boing Boing's story on trying and failing to manufacture Monopoly in the U.S. reads as a parable but is a supply chain audit in miniature. The company ultimately cancelled the U.S. expansion because tariff costs exceeded the production advantage of onshoring. That's the structural argument against tariff-driven reshoring in a single product: when the components don't exist domestically, the tariff doesn't create a domestic supplier — it creates a margin problem. For any brand doing reshoring math right now, the dice story is worth keeping in the deck.

15 articles across 6 themes · 12 sources · Powered by Folo + Claude