The Adjacent Brief

TL;DR: China's government ordered its tech platforms to stop competing on price and redirect capital toward AI investment, while Apple's eyewear ambitions drew fresh comparisons to its 2015 watch playbook. A federal lawsuit between Amazon and Perplexity may settle whether AI agents have legal standing to crawl the web at all.

Worth Reading

• Uber and Lyft drivers keep working despite chronic complaints — here's the actual reason — Flexibility plus low entry cost beats every available alternative — which tells you something about the labor market underneath the gig economy.
• User-replaceable batteries are returning — and EU legislation is the reason — Regulatory design outpacing market design, one Right to Repair win at a time.
• Samsung's Galaxy S Ultra needs a strategic rethink — The premium tier is losing the argument when the standard flagship does 90% of the work at 70% of the price.
• Is AI actually more expensive than the employees it's replacing? — Scott Galloway and Ed Elson run the numbers; the answer depends heavily on what you're measuring and over what horizon.
• Pangram claims a 1-in-10,000 false-positive rate — that's not reassuring at scale — The Atlantic's look at AI detection's most trusted tool finds that "gold standard" accuracy at population scale still produces thousands of false accusations.
• Young moviegoers pushed 'Backrooms' to $82 million at the box office (paywall) — Internet-native IP crossing into theatrical is a story the film industry keeps underweighting.

Brand & Growth

The price war ends by decree

China's government published a directive in Qiushi — the Communist Party's flagship policy journal — ordering major internet platforms to stop competing on price and start investing in AI. For Alibaba, JD, Meituan, and Pinduoduo, years of subsidy warfare and margin-burning promotions had become the competitive default. Beijing is calling time on that model, and the mechanism is explicit: capital that was going into discounts should go into models, chips, and infrastructure. This is less a technology policy than a capital reallocation mandate. China's consumer platforms will look more like R&D vehicles and less like retail operations, giving state-backed AI development a funding tailwind paid for by the platforms themselves.

The watch playbook, reprinted for eyewear

Apple is preparing a move into the $200 billion eyewear market, and the framing in Apple's eyewear ambitions is deliberate: this is the Apple Watch story replayed. In 2015, Apple entered a fragmented mid-market watch category where Swiss brands owned the top and cheap quartz owned the bottom. Within three years it was the world's bestselling watch. Eyewear is structurally similar — Luxottica controls the mid-to-premium frame market, margins are high, and the category has seen almost no technology disruption. Smart glasses from Meta and others have chipped at awareness without cracking the mainstream. Whether Apple replicates the watch outcome depends on form factor and whether the use case is compelling enough at launch to justify wearing a computer on your face daily, a bar Vision Pro demonstrably did not clear.

Disney consolidates, Hulu disappears into the bundle

Disney's plan to absorb Hulu into Disney+ as a standalone app ends as the logical conclusion of a strategy that began when Disney acquired full Hulu control in 2023. The standalone app held Hulu's subscriber base while Disney built out Disney+ content depth. With that library now larger, maintaining two apps means maintaining two acquisition funnels, two support infrastructures, and two brand positions that increasingly overlap. One app with tiered content access is simpler to sell, simpler to retain, and simpler to bundle with ESPN+. The casualty is Hulu's distinct brand identity, which had real equity among cord-cutters who saw it as the less Disney-fied option.

Connected World

Hardware costs are the variable nobody priced in

The post-pandemic assumption was that component costs would normalize. They haven't, and the combination of rocket failures, supply chain disruption, and component price inflation is compressing margins across consumer hardware in ways manufacturers didn't model for. Steam Deck is the case study in The Register's piece — a device that launched at a competitive price point and now faces pressure from both ends: rising BOM costs and a consumer trained to expect price drops, not increases. Any hardware product line launched on 2022 cost assumptions is running a margin deficit that either gets passed to consumers or absorbed until it can't be.

Nvidia moves up the stack

Nvidia is shipping ARM-based CPUs optimized for local AI agent inference, with CUDA support baked in from the start. The timing is more notable than the move itself. Nvidia has spent three years as the indispensable infrastructure layer for cloud AI, and stepping into the CPU business extends that position to the full inference stack. Local AI inference is where enterprise AI actually lands: closer to the data, cheaper per query at scale, and outside the latency and cost constraints of API calls. If Nvidia owns both the training infrastructure and the deployment endpoint, it controls the economics of AI at every layer.

The off-grid edge

At the smaller end of the infrastructure story: an iPhone-powered off-grid OCR server built by a hobbyist is doing real document processing work without any cloud dependency. It's a Hackaday project, not a product launch. On-device AI inference has matured to the point where a consumer smartphone can run a useful, production-adjacent service in isolation — the same trend Nvidia is betting on at enterprise scale, visible here as a weekend build.

Culture & Signal

Western AI tools are now Iranian cyberweapons

ChatGPT and Gemini are being used operationally by Iranian cyber units to accelerate malware development and run more sophisticated phishing operations, according to a Financial Times report citing experts across multiple security firms. This is documented operational reality. The policy implication is uncomfortable: the same openness that makes these models commercially valuable makes them available to adversaries at zero marginal cost. Export controls on chips don't extend to API access, and usage monitoring at the individual query level is unsolved. For enterprise security teams, the practical question is what happens when a phishing email is indistinguishable from legitimate correspondence because it was drafted by the same model their employees use.

The faith model

LLMs as a form of institutional religion is the argument in The Register, and it's more precise than it sounds. LLMs share the structural features that make religious institutions durable and dangerous: answers delivered with authority regardless of epistemic grounds, intermediary figures who claim special interpretive access, and communities that treat skepticism as apostasy. The practical warning is about the people who benefit from that dynamic — vendors, influencers, and consultants whose revenue depends on AI being perceived as oracular rather than probabilistic. The Vatican's actual AI encyclical this month, which reportedly gives Catholic employees potential legal grounds to refuse workplace AI systems, is a reminder that these institutional dynamics run in both directions.

A lawsuit that rewrites the rules for web access

Amazon's CFAA case against Perplexity is narrower than most coverage suggests — it turns on whether Perplexity's AI agent scraping constitutes unauthorized computer access under the Computer Fraud and Abuse Act — but the precedent it could set is not narrow at all. If the court finds that AI agents violate the CFAA when they crawl sites without explicit permission, the legal infrastructure of the open web changes. Every RAG pipeline, every agentic workflow that reads external sites, every AI product that pulls live content would require affirmative access agreements. The party that benefits most from a ruling like that is Amazon, which has the resources to negotiate those agreements at scale and the existing relationships to make them quickly.

The New Consumer

Detection tools create new categories of collateral damage

Pangram, marketed as the most accurate AI writing detector available, claims a false-positive rate of 1 in 10,000. The Atlantic's analysis finds that number is real — and entirely inadequate. At a large university processing 100,000 submissions per semester, that rate produces ten wrongful academic misconduct flags per term. Across a national educational system, it produces thousands. The product is being deployed in contexts where the base rate of the thing it's detecting, and the cost of a false accusation, make even very low error rates consequential despite meeting normal software standards. Any institution using AI detection tools at scale without a formal appeals process and error-rate disclosure is carrying liability it hasn't priced.

Craft flour as cultural signal

Premium flour sales are rising (paywall) sits at the intersection of two otherwise unrelated consumer movements: the MAHA (Make America Healthy Again) influencer circuit and the chef-driven whole-grain revival. The NYT piece documents mills that were marginal operations three years ago now running waitlists. The editorial interest is the coalition — a demographic that follows RFK Jr.-adjacent wellness content and a demographic that follows Eater are buying the same product for different reasons. That kind of dual-market convergence typically precedes a category going mainstream, which is when the big CPG players start making acquisitions.

Internet-native IP is box office now

Backrooms crossed $82 million in ticket sales (paywall), powered by young audiences who came in with pre-existing lore fluency. A24 backed a film based on a creepypasta genre with no Hollywood IP lineage, and it outperformed most franchise sequels this spring. The mechanism matters: audiences who already know the fictional universe don't need the film to do worldbuilding, which compresses runtime, reduces exposition, and lets the film go somewhere. For studios deciding where to invest development resources, this is behavioral data — not a survey about what Gen Z says it wants to watch — showing that internet-native IP has genuine theatrical traction when produced with craft.

Machines & Minds

The ROI gap is structural, not a rounding error

AI adoption follows a J-curve, Exponential View argues, drawing the comparison to early US factory electrification — a period when firms spent decades rewiring physical infrastructure before productivity gains showed up in output statistics. The analogy is sobering for anyone building quarterly ROI cases for AI investment. The productivity gains from electrification weren't visible in aggregate data for roughly 20 years after adoption began, because they required reorganizing work around the new capability, not just plugging in motors. The same dynamic applies here: deploying a model into an existing workflow produces marginal gains. Restructuring the workflow around model capabilities produces step-change gains, and that restructuring takes time, management attention, and often personnel changes that most organizations aren't ready to make.

Data quality as the actual competitive moat

The returns to good data are rising, Tyler Cowen argues at Marginal Revolution, and the argument connects directly to the electrification piece. If productivity gains from AI depend on workflow restructuring, then the organizations that restructure fastest are the ones with clean, queryable, well-labeled internal data. Government data is a particular bottleneck: poorly structured public datasets limit the entire class of AI applications built on public-sector information. The commercial implication is that data cleaning and governance tooling — an unglamorous category — may have better near-term returns than the AI application layer it enables.

Token economics are restructuring AI pricing

The token reckoning Big Technology describes is a business model crisis. As token prices drop (Sam Altman has floated the idea that AI intelligence costs will eventually converge toward the cost of electricity), revenue models built on per-token pricing compress. Vendors who priced on token volume need to migrate to value-based pricing: outcomes, workflows, seats. This is what enterprise software learned going from on-premise licensing to SaaS, and the transition is neither smooth nor fast. The firms that get there first — by demonstrating measurable output value rather than selling compute — will own the enterprise contracts when the commodity price floor hits.

"Human in the loop" is a liability, not a safeguard

Human oversight of agentic AI is not working as designed, SiliconAngle reports — because the oversight mechanisms in place are cosmetic. Humans positioned as checkpoints in autonomous workflows are approving outputs they don't have the context, time, or tooling to evaluate. The result is that "human in the loop" functions as liability coverage rather than actual control. The piece is careful not to prescribe a fix, which is appropriate — meaningful human oversight of complex agentic tasks requires redesigning the interface, not just inserting an approval button. For enterprise AI buyers, this is the question to ask vendors before signing: what does your oversight mechanism actually let a human catch?

Commerce Rewired

The funding rebound has a composition problem

Black founders raised more venture capital in Q1 2026 than any quarter since 2022, according to TechCrunch — but the catch TechCrunch names is worth reading carefully. The headline number is real; the distribution is not. A small number of late-stage rounds, in companies that were already established before the 2022 capital pullback, account for the majority of the dollar total. Seed and Series A activity — the stages where structural access gaps show up most acutely — have not recovered at the same rate. A funding rebound concentrated at late stage leaves the pipeline question unanswered: the cohort of new companies that would generate the next round of late-stage deals is not forming at the pace the headline suggests.

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