The Adjacent Brief

TL;DR: AI's constraint has shifted from capability to consequence — the systems are now powerful enough to enter domains where being wrong has real costs, and the trust infrastructure to govern them doesn't exist yet. Anthropic is rattling banks, Meta is dispensing bad health advice from raw biometric data, and the global AI arms race is accelerating with no institutional brakes. Meanwhile, the tools are making their users shallower thinkers. Power without accountability isn't a technical problem — it's a structural one, and no one has solved it.

Worth Reading

• AI enters high-stakes domains before trust infrastructure exists — Anthropic's latest model isn't impressive because it's capable. It's alarming because banks can't tell when it's wrong.
• The global AI arms race has no off switch — The U.S., China, and Russia are each building AI weapons systems while assuming the others won't stop. None of them will.
• Meta asked for health data and gave dangerous advice in return — The most important data point about AI health apps isn't the feature. It's what happens when the model gets it wrong.
• AI screens 15 million molecules a day and still can't cure Alzheimer's — Speed without interpretability is just faster failure. Drug discovery's ceiling isn't compute.
• Siri is being left behind — and Apple has no one to blame but itself — Apple spent a decade treating Siri as a feature, not a product. Now it's paying the consolidation tax.
• AI is making us faster and shallower at the same time — Productivity gains are real. So is the cognitive atrophy. Organizations optimizing for the former are accumulating the latter.
• Berlin fintech Credibur scales to €2B in debt volumes in months — The embedded lending infrastructure layer is being built out faster than most incumbents realize.

Machines & Minds

Powerful enough to do damage, not accountable enough to stop it

Anthropic's Claude 3.7 is striking fear into bank compliance and legal teams not because it's threatening jobs but because it's capable enough for consequential decisions — and no one has figured out what happens when it fails at scale. The liability frameworks, audit trails, and failure-mode documentation that financial institutions require don't exist in any meaningful form. That gap is the product risk, not the benchmark score.

Meta's AI health feature follows the same pattern with fewer guardrails and a consumer audience. The model solicited raw biometric data — heart rate, sleep, symptoms — then returned medically irresponsible advice. This isn't a hallucination story. It's a deployment decision story: someone at Meta chose to ship a health feature before building the epistemic scaffolding that health applications require. The trust debt is being incurred now. The bill arrives when someone acts on bad guidance.

Speed is not the same as progress

Generative AI has genuinely transformed drug discovery throughput — screening 15 million molecules in a day is not trivial. But the ceiling isn't processing speed. It's interpretability and target selection: the models can rank candidates but can't explain why a candidate is promising in a way that guides clinical hypotheses. Alzheimer's remains unsolved not because we lack molecule candidates but because we don't understand the disease mechanism well enough to know what we're selecting for. AI can accelerate a search through a poorly defined space — it can't define the space for you.

The arms race with no rulebook

The escalating global AI weapons buildup — the U.S., China, and Russia each developing autonomous targeting, AI-assisted command systems, and algorithmic threat response — is the trust problem at geopolitical scale. Each actor is accelerating because it assumes the others won't stop. There's no verification mechanism, no treaty framework with teeth, and no shared definition of what "AI-enabled weapons" even means across jurisdictions. This dynamic is now self-reinforcing in ways that diplomatic timelines can't match.

The tools are outrunning the people using them

Gary Marcus's evaluation of Claude Mythos and The Next Web's analysis of AI and cognitive atrophy converge on the same finding: productivity is going up and cognitive depth is going down. Workers using AI assistance are completing more tasks but building fewer transferable mental models. Organizations that optimize for throughput are accumulating reasoning debt they won't see until a novel problem arrives — one the AI can't pattern-match and the human can't think through. The same dynamic is playing out in creative work: the ethics of AI-assisted creative output aren't just about disclosure — they're about what happens to craft when the friction of making is removed.

Siri is the cautionary tale, not the comeback story

Apple's Siri problem, reported in depth this week, is structural. Apple spent years treating Siri as a gateway to device features — a voice UI for calendar entries and timer settings — while OpenAI, Anthropic, and Google built reasoning systems. The gap isn't fixable with a model update. It reflects a decade of product decisions that treated the assistant as peripheral rather than platform. Apple is now paying the consolidation tax: in a world where one or two AI assistants become default infrastructure, incumbents who didn't invest when it was optional will fight for scraps.

What gets buried kills you later

Phil McKinney's analysis of ghost research inside R&D organizations — work that exists in institutional memory but was never formally completed, documented, or killed — matters more in an AI context than the framing suggests. As AI tools generate research outputs faster than organizations can evaluate them, the ghost research problem will compound. Teams will have more findings, less closure, and no process for knowing what was validated. Speed makes the institutional blind spot wider, not smaller.

Connected World

Your face is the next contested data asset

[Spyglass's Dossier #003 on the AI battle for biometric face data maps a pattern that's accelerating: facial recognition capability is now commodity infrastructure, which means the competitive moat has shifted to data exclusivity and model training rights](

https://www.morningbrew.com/stories/2026/04/11/anthropic-latest-ai-model-strikes-fear-into-banks?utm_source=&utm_medium=syndication&utm_campaign=feed

). The question isn't whether facial AI exists — it's who owns the datasets that make it accurate across demographics, and what the rights framework is for the people whose faces trained the model. No country has resolved this. Several are actively avoiding it.

Capital is starting to back the climate infrastructure layer

The Israeli carbon capture startup Repair securing Shell backing and scaling of European operations is a modest data point in a larger pattern: the infrastructure layer beneath net-zero commitments is finally attracting institutional capital. Shell's involvement matters less as a climate signal than as a strategic one — major energy incumbents are hedging against stranded assets by funding the abatement technologies their own operations will eventually need to buy. That's a different bet than a climate VC making a thesis investment.

The New Consumer

Young audiences routing around institutional gatekeepers — results vary

Gen Z's growing financial sophistication — documented in Morning Brew's look at how younger investors developed market literacy — is real, but its distribution is uneven in ways that aggregate data obscures. The cohort that discovered index funds on TikTok is meaningfully different from the cohort that discovered leverage on TikTok. The same disintermediation that created self-directed investors also created the scholarship fraud ecosystem that NYT documented this week: fake scholarship listings on social media platforms targeting families who distrust the financial aid system enough to bypass it. Routing around gatekeepers works when the routing is reliable. When the alternative channel is unregulated, it produces predation at scale.

Manifestation goes mainstream, and the tell is the branding

Casey Lewis's close read of disco manifestation and Scientology-adjacent wellness content identifies something worth watching: the wellness-to-belief pipeline has accelerated. When manifestation content moves from niche TikTok to branded product drops and live events, it stops being subculture and becomes institutional structure with economic dependencies. The tell isn't the content — it's the monetization architecture underneath it. When a belief system develops a commerce layer, it has incentives to deepen commitment rather than improve outcomes. Brands building in wellness-adjacent space should map where their audiences sit on that pipeline.

Meta's cafeteria workers beat ICE — and the story is about corporate leverage, not immigration

The Meta cafeteria workers who successfully resisted an ICE operation on campus succeeded because Meta's facilities management structure created a leverage point that individual workers wouldn't have had in isolation. The story gets filed under labor and immigration, but the operative mechanism is corporate institutional power being activated — whether deliberately or by circumstance — in ways that override federal action. That's a pattern worth watching as enforcement activity escalates. The question of where large private employers land on these confrontations is no longer theoretical.

Commerce Rewired

Embedded lending is building its infrastructure layer

Credibur reaching €2B in debt volumes within months of scaling is a data point in a pattern the embedded finance space has been building toward for several years: the B2B debt infrastructure layer — the rails that let non-bank platforms originate, structure, and service credit — is now scaling fast enough to show up in headline numbers. The signal isn't the €2B itself. It's that an infrastructure-layer fintech can reach that volume without being a consumer brand, a marketplace, or a lending institution in the traditional sense. The middle layer of finance is being rebuilt by companies most people have never heard of.

The short-term rental market is repricing around accessibility

AirDNA naming the Finger Lakes the best budget-friendly short-term rental market of 2026 matters less as a travel tip than as a signal about where demand is migrating. The premium coastal and mountain markets saturated; the arbitrage has shifted to secondary markets with sub-$250K acquisition costs, drive-to accessibility, and underbuilt hospitality infrastructure. For investors and operators, the playbook has moved from "find a desirable location" to "find a location that's about to become desirable because the desirable ones got too expensive."

19 articles across 6 themes · 14 sources · Powered by Folo + Claude