> ## Content Index
> Fetch the complete content index at: https://adjacent.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Adjacent Brief — August 24, 2026
- URL: https://adjacent.media/briefs/2026-08-24/
- Published: 2026-08-24T14:15:02.000Z
- Updated: 2026-08-24T14:15:03.000Z
- Description: Harvard Business School is selling a $699 bootcamp taught by AI-generated avatars of its own faculty. A proposed class action against Oura argues the ring’s sleep-stage readings are model estimates rather than measurements.
- Author: Jonathan Greene
- Tags: #brief

**TL;DR:** Harvard Business School is selling a $699 bootcamp taught by AI-generated avatars of its own faculty. A proposed class action against Oura argues the ring's sleep-stage readings are model estimates rather than measurements. A stealth model called Ox Alpha has climbed OpenRouter's developer rankings with no disclosed owner or infrastructure. Cost and provenance questions run through each item.

## Worth Reading

- [Harvard will sell you a $699 course taught by clones of its own professors](https://www.nytimes.com/2026/08/22/business/dealbook/harvard-ai-faculty.html?ref=adjacent.media) (paywall) — The crest is the product; the faculty are the render layer.
- [Chinese models are closing the capability gap while undercutting on price](https://www.bloomberg.com/graphics/2026-us-china-ai-race/?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc4NzM3MzM3OSwiZXhwIjoxNzg3OTc4MTc5LCJhcnRpY2xlSWQiOiJUSzFCTzFLR1pBSVowMCIsImJjb25uZWN0SWQiOiIwNEFGQkMxQkYyMTA0NUVEODg3MzQxQkQwQzIyNzRBMCJ9.fkDLdJ6qkTKX0FAK%5FsbPEYQcSfeBcZcdbIQkeVOkzMM&ref=adjacent.media) (paywall) — Bloomberg's charts are the buyer's case for keeping a second provider wired in.
- [A billion dollars of training spend buys a model nobody can tell apart](https://www.gettheleverage.com/p/a-billion-dollars-buys-you-nothing?ref=adjacent.media) — If capability converges, the moat has to be distribution or workflow.
- [Oura's sleep stages are estimates, and a lawsuit wants them labeled that way](https://thenextweb.com/news/oura-sleep-tracking-accuracy-class-action-lawsuit?ref=adjacent.media) — The first real test of what "measurement" means when inference is doing the work.
- [Amazon is cutting up rare books to get clean pre-2022 training text](https://www.404media.co/podcast-amazon-is-destroying-rare-books-to-train-ai/?ref=adjacent.media) — Uncontaminated human writing has become a physical supply chain.
- [Delivery delay is the feature: slow-texting apps are pulling real users](https://www.nytimes.com/2026/08/22/technology/carrier-pigeon-app-texting-technology.html?unlocked%5Farticle%5Fcode=1.7VA.5e54.GDSmjwFMUqzc&&ref=adjacent.media#x26;smid=url-share) — Carrier Pigeon at 75K users, Roost at 650K downloads; also carried \[in the business section\](https://www.nytimes.com/2026/08/22/technology/carrier-pigeon-app-texting-technology.html).

## Connected World

**Electrification's disposal cost lands on whoever runs the bin lorry**

Lithium-ion batteries hidden in household waste are [costing the UK roughly £1bn a year in recycling-centre fires](https://hardware.slashdot.org/story/26/08/23/0027208/battery-fires-at-recycling-centres-are-costing-the-uk-1bn-a-year?utm%5Fsource=rss1.0mainlinkanon&utm%5Fmedium=feed) — vapes, power banks, e-bike packs, all crushed into a waste stream never designed to hold energy storage. The number matters because of who's paying it: councils, waste contractors, and their insurers, none of whom sold the battery. That's the classic setup for extended producer responsibility rules, and any brand shipping a cell inside a low-cost consumer good should assume the take-back cost gets pushed back up the chain within two budget cycles. The second-life battery market we've been tracking gets its economics from exactly this pressure: reuse looks expensive until disposal is priced properly.

**China is siting compute by decree while US governors reverse themselves**

Ulanqab, a city in Inner Mongolia with cheap coal and wind power, spare land, and a short fiber run to Beijing, now hosts [around 100 data centers built or under construction](https://www.wired.com/story/the-unlikely-place-at-the-center-of-chinas-ai-boom/?ref=adjacent.media). Wired's reporting describes a placement decision made once, centrally, and executed. Compare that with the US, where Abbott and Shapiro have both walked back data-center approvals they previously championed after constituent opposition made siting a political liability. China isn't winning outright here — inland siting carries latency and staffing costs. But land and power access are becoming the binding constraint on buildout schedules, and that constraint resolves differently under different political systems. The same buildout is squeezing printed circuit boards and multilayer ceramic capacitors hard enough to raise prices threefold for automotive buyers, a thread worth watching if you procure anything with a board in it.

**The humanoid category's biggest vendor is talking down his own market**

At the World Robot Conference in Beijing, with [over 300 exhibitors on the floor](https://www.ft.com/content/e16ded89-b618-4952-a0ab-96ef11d06582?accessToken=zwAAAaAngdVAkdPhbe2JthhJUtOgq5bvEdBlgg.MEUCIEVQhJrsMuaQJtWVDllL5U7XIUfljfJ9TsHat-OofHhlAiEA3OJBJeuvk038EzzRnTfhdtFbwtS5Xi50FK491b6L-S0&&ref=adjacent.media#x26;sharetype=gift&token=76223979-4b23-4ab0-a4a7-1b2d5f74c331) (paywall), Unitree founder Wang Xingxing told the FT the industry's "ChatGPT moment" hasn't happened yet. Founders raising capital do not usually deflate their own category, which makes the comment more useful than the demo reels around it. Read it as a schedule: humanoids remain pilot-stage hardware sold into logistics and inspection trials, not a line item for 2027 labor planning.

## The New Consumer

**Friday night got redistributed, and nobody adjusted their staffing model**

Time-use data on [the collapse of the Friday-night gathering](https://marginalrevolution.com/marginalrevolution/2026/08/the-end-of-friday-nights-with-friends.html?utm%5Fsource=rss&utm%5Fmedium=rss&utm%5Fcampaign=the-end-of-friday-nights-with-friends) is the kind of behavioral evidence the loneliness discourse usually lacks: this is what people did, not what they told a pollster they felt. Socializing hasn't vanished so much as spread thinner across the week and shifted earlier, with the sharpest changes among younger adults. If you operate restaurants, cinemas, bars, or anything with a Friday peak baked into labor scheduling and promotional calendars, the demand curve underneath those assumptions has flattened. In Bed With Social's read on [what actually broke through this summer](https://open.substack.com/pub/maried/p/this-summers-biggest-hit) points at the same fragmentation from the culture side: the hits are still real, but the shared appointment window they used to arrive through is gone.

**A lawsuit is about to define what "measurement" means on a spec sheet**

The proposed class action arguing that [Oura's sleep stages are model estimates with roughly coin-flip accuracy](https://thenextweb.com/news/oura-sleep-tracking-accuracy-class-action-lawsuit?ref=adjacent.media) is more consequential than a single wearable's marketing copy. Every consumer health device infers something it cannot directly observe; the ring's PPG sensor and accelerometer are not a polysomnogram, and the industry has been comfortable letting a confident number stand in for a probability. What the suit tests is whether presenting inference as reading is a design choice or a claim. For any product team shipping AI-derived outputs to consumers, the practical move is immediate: confidence intervals in the UI, methodology in the help docs, and legal review of every verb in the app store listing that implies detection rather than estimation.

## Machines & Minds

**Convergence at the frontier means the model stopped being the asset**

The argument that [a billion dollars in training spend now yields nothing a customer can distinguish](https://www.gettheleverage.com/p/a-billion-dollars-buys-you-nothing?ref=adjacent.media) gets empirical backup from two directions. Bloomberg's analysis of [the narrowing US–China gap](https://www.bloomberg.com/graphics/2026-us-china-ai-race/?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc4NzM3MzM3OSwiZXhwIjoxNzg3OTc4MTc5LCJhcnRpY2xlSWQiOiJUSzFCTzFLR1pBSVowMCIsImJjb25uZWN0SWQiOiIwNEFGQkMxQkYyMTA0NUVEODg3MzQxQkQwQzIyNzRBMCJ9.fkDLdJ6qkTKX0FAK%5FsbPEYQcSfeBcZcdbIQkeVOkzMM&ref=adjacent.media) (paywall) documents Chinese labs shipping releases that clear enterprise capability thresholds at prices US labs can't match. On OpenRouter, a model called Ox Alpha [climbed the developer rankings with no disclosed owner and no known infrastructure](https://thenextweb.com/news/ox-alpha-stealth-model-openrouter-anonymous-provider?ref=adjacent.media) behind it. That last one is the tell: developers routed real traffic to a model whose provenance they could not verify, because the benchmark scores and the price were good. Brand equity in frontier models is thinner than the marketing budgets suggest.

**Agent systems fail like O-rings, and that's a pricing problem**

Kremer's O-ring model — output as the product of every task's quality, so the weakest step drags the whole chain — describes [what agentic workflows actually do in production](https://marginalrevolution.com/marginalrevolution/2026/08/the-new-agentic-o-ring-world.html?utm%5Fsource=rss&utm%5Fmedium=rss&utm%5Fcampaign=the-new-agentic-o-ring-world). Chain ten steps at 95% reliability and you ship a 60% system. That math explains why agents still need constant human context injection and why founders describe supervising them as a full-time job rather than a saved one. It also raises the value of whoever holds the last checkpoint. The governance version of the same gap shows up in TechCrunch's finding that [frontier labs still won't describe how they'd contain a model that misbehaves](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/?ref=adjacent.media) — not a refusal so much as an absence, which is what makes it awkward for enterprise risk committees being asked to sign agentic deployments. Base purchasing decisions on measured reliability across the full chain rather than on per-step benchmarks.

## Commerce Rewired

**Virtual try-on works because returns are a measurable cost, and fit is a solvable problem**

Zalando, Zara, and ASOS are deploying [AI fitting rooms aimed squarely at return rates](https://www.bloomberg.com/news/articles/2026-08-21/zalando-zara-use-ai-virtual-try-ons-to-tackle-clothing-returns?ref=adjacent.media) (paywall), and that framing is why this deployment deserves more attention than most retail AI. Online apparel returns run 25–40% in Europe, and each one carries shipping both ways, inspection, repackaging, and frequent markdown or write-off. A tool that shaves even a few points off that has a payback period a CFO can calculate — a different thing entirely from the engagement-metric AR try-ons retailers shipped five years ago and quietly retired. The lesson generalizes: the AI features surviving 2026 budget review are the ones measured in reverse logistics, contact-center deflection, and write-offs avoided.

**Token pricing is a lock-in mechanism wearing a utility-billing costume**

SiliconANGLE's argument that [whoever defines the token is defining your margin](https://siliconangle.com/2026/08/22/from-tokenmaxxing-to-sovereign-alpha-who-controls-your-ai-economics/?ref=adjacent.media) lands against the convergence story above. Vendors are standardizing on tokens and API calls precisely because those units are non-portable: a per-token price can't be benchmarked cleanly across providers when tokenizers, context handling, and caching all differ. That opacity is the switching cost, and it exists because the underlying models are getting harder to differentiate. Practical response for anyone signing an enterprise AI contract this quarter: negotiate on cost per completed task or per resolved ticket, insist on usage telemetry you own, and keep a second provider integrated even if you never route to it.

## Brand & Growth

**Harvard is monetizing the crest, and the faculty are the rendering**

The $699 bootcamp [taught by AI clones of Harvard Business School professors](https://www.nytimes.com/2026/08/22/business/dealbook/harvard-ai-faculty.html?ref=adjacent.media) (paywall) is a clean piece of brand arithmetic: near-zero marginal delivery cost, unlimited seats, and a price point that clears without touching the MBA's scarcity. The crest is what's being sold; the avatars are a distribution mechanism for it. Two risks sit under that. The first is dilution: every dollar of low-cost credential revenue tests how much the premium credential's value depends on being hard to get. The second is that the faculty likeness becomes a licensable asset with its own negotiation, a labor question universities have not begun to price. For anyone running a brand with expensive scarce human delivery — consultancies, law firms, agencies, medical practices — this is the template being tested in public. Renewal and completion rates reveal whether a course actually teaches; enrollment only measures the brand.

---

*14 articles across 5 themes · 11 sources · Powered by Folo + Claude*