Apple's first foldable iPhone now has a product page

Top Story

Apple’s first foldable iPhone now has a product page, which means the foldable has stopped being a category you opt into and started being one you default into. iPhone Duo: a 7.6-inch inner Super Retina XDR display, titanium frame and hinge cover, vapor-cooled A20 Pro, a dual-battery system, an under-display FaceTime camera, and iOS 27 reworked around the shape. Seven years after Samsung shipped the original Galaxy Fold and roughly five after everyone in the industry settled on “Apple is two years away,” the company has put its flagship behind a hinge.

The engineering claims worth taking seriously are the ones about the display stack. Ten ultrathin layers, a nano-texture finish to cut glare, 3000 nits peak, wide-angle OLEDs, and a camera hidden beneath the panel. That last one is the tell. Foldables have shipped with visible punch-holes and off-center selfie cameras since 2019 precisely because burying a sensor under a folding OLED without wrecking the panel is hard. Apple claims it solved it, which would make the Duo’s inner display the first foldable panel designed without an exception carved into it.

Slow down at the numbers. Apple says the Duo’s display is “50 percent larger than iPhone 18 Pro Max.” A 7.6-inch diagonal against a Pro Max-class panel of roughly 6.9 inches is about 21 percent more area at matched aspect ratio; reaching 50 percent takes about 8.4 inches. Either Apple is measuring usable app area in a specific pose, or the comparison isn’t against what the sentence implies. The company also hasn’t published price, ship date, weight, ingress rating, or hinge cycle rating. Those absences are the real spec sheet, and until a teardown fills them in, “50 percent” is marketing copy rather than a measurement.

The other figure worth flagging — the outer display at “more than 90 percent of the screen area of iPhone 18 Pro” — is the one competitors should worry about. A cover panel nearly the size of a whole phone is a different engineering trade than a slim notification strip, and it is what makes the device usable closed rather than something you constantly open. If that holds, the “you’re carrying a small phone when folded” objection that has kept the category in the low single-digit millions is largely answered on paper.

The weakest line on the page is “Meet Siri AI,” the marquee feature of the new flagship, footnoted to “rolling out in English.” Apple’s assistant has slipped enough times that shipping the Duo alongside a rolling-out assistant restates Apple’s standing bet: the phone is the interface, and someone else’s model is the substance. That landed the same day a pretraining researcher resigned from a frontier lab over safety, DeepSeek undercut its own v4 Pro with a cheaper, more capable v4.1 Flash, and Anthropic was reported to be building a predictive surveillance system aimed at activists. Apple’s answer to all three is to own the handset and let the models commoditize.

Foldables stalled for hardware-adjacent reasons but mostly software ones. Google’s problem was never the hinge; it was that Android apps treat the inner screen as an afterthought. Apple controls the app environment outright, and the posable-form-factor work in iOS 27 — StandBy without a charger, the reworked conversation window — is a software pitch rather than a hardware one. Samsung, which has carried this category alone since 2019, now has to compete on an ecosystem it doesn’t own.

Watch the price and the first hinge teardown. If the Duo lands above the Pro Max, the foldable stays a halo product for another cycle. If Apple prices it at parity, a seven-year head start is worth less than the quarter it takes to lose it. iPhone Duo

Also Today

Apple’s New iPhone Camera Mode Promises to Prove Your Photo Isn’t AI · Source Apple’s Reference Image mode signs every pixel the iPhone 18 Pro’s sensor captures, develops it in Private Cloud Compute into what Apple calls an unalterable reference image, and lets you diff that against later edits. A Reference Image API reaches iOS, iPadOS, and macOS; capture misses the EU at launch, though EU users can view reference images on next year’s OS versions. Hardware-rooted signing is strictly stronger than C2PA-style metadata, which survives only until something re-encodes the file. It also proves less than the framing suggests — that the frame came off the sensor unchanged, not that the scene in front of the lens was real, and only in one camera mode. As provenance standards go, this is the first with a plausible enforcement mechanism, which matters more than how much it proves.

I Resigned from Anthropic Today · Source A pretraining researcher with three years across OpenAI and Anthropic resigned publicly on September 9, writing that neither lab is acting responsibly and that both are “racing straight to self-improving superintelligence and gambling with our lives.” The substance isn’t the accusation but the mechanism: Anthropic’s staff reportedly understand the stakes and build anyway, because they assume nobody else will stop first. The concrete ask is pacing agreements between US labs and a possible temporary ban on capability improvements, which the author argues the Hugging Face attack made more viable. One reply attributes to Anthropic alignment lead Evan Hubinger a >10% personal extinction estimate and no plan for safely aligning superintelligence — unverified here, but the claim worth chasing. If the people who understand the systems best keep leaving with the same paragraph, the labs have a retention problem that no safety framework document fixes.

DeepSeek launching v4.1 flash cheaper and more capable than v4 pro · Source DeepSeek says V4.1 Flash, due around September 10 Beijing time, beats V4 Pro on performance, cost, speed, and completion time — and that it will route all V4 Pro requests to Flash, billed at Flash’s price, until V4.1 Pro ships. Off-peak Flash pricing lands at $0.003 per million input cache hits, $0.15 cache misses, $0.6 output, doubled at peak. The Hacker News objection is the right one: a model name is a contract, and silently swapping the weights underneath it invalidates whatever eval work a customer did against the old distribution, however much better the new model benches. DeepSeek is undercutting its own flagship, which tells you the margin floor sits below what the price list implied. Anyone with a pinned model name in production should read this as a warning about the whole hosted-API model, not about DeepSeek specifically.

Tailwind Labs is joining Shopify · Source Tailwind is joining Shopify. The framework keeps its MIT license and its team; what ends is the business around it — Tailwind Plus and ui.sh stop taking new customers, and 110 million weekly installs become a cost Shopify carries rather than a product anybody sells. Adam Wathan’s stated reason is that he always wanted the framework developed in service of a real application, and Shopify supplies one: storefronts, admin, checkout, agentic commerce. That’s a better development loop than a template business and a narrower one, since the roadmap now bends toward a single merchant platform’s problems. The web’s most-used utility framework just stopped being independent and became a platform detail of one large retailer — stable, probably well-maintained, and no longer accountable to anyone else.

Google to Invest Record $15 Billion In AI Infrastructure In Finland · Source Google is putting €13 billion into Finnish digital infrastructure over two years, its largest single European investment: expansion at Hamina plus a portfolio of power deals. The power is the interesting part. Google signed a 22-year agreement supporting life extension of the Loviisa nuclear plant, added onshore wind, and contracted a 94 MW battery to hold prices through cold, windless periods — firm capacity, not certificates. Google claims 37,000 jobs supported during 2027–28 construction, €3.6 billion a year in GDP contribution, and €31 million for local colleges. The company is buying the input AI buildouts actually run short of. Nuclear life extension and grid batteries are the tell: European AI capacity is now a grid-access problem.

In Brief

• Automattic CEO Matt Mullenweg is on a leave of absence, announced on a company-wide Slack this morning, with no duration or interim leadership reported. (Source)
• Anthropic job postings and interviews with its senior security staff describe an extensive monitoring system aimed at activists who oppose rapid AI development — in the same week one of its own pretraining researchers resigned over the pace. (Source)
• Waymo rides can now be booked through the Lyft app in Nashville, another step toward a market where the brand on the app and the software driving the car are different companies. (Source)
• Four women are suing Amazon, alleging the company denied basic accommodations such as bathroom breaks to pregnant warehouse workers and then fired them. (Source)
• OpenAI published a claimed solution to the Navier–Stokes Millennium Prize problem, which is remarkable enough that independent verification, not the blog post, is the news. (Source)
• A roundup argues that OpenAI’s GPT-6 Astra rests on looped, recurrent-depth transformers, an architecture choice that would explain both its cost profile and its “hidden reasoning” — internals worth treating as inference until OpenAI confirms anything. (Source)
• Alibaba’s Qwen3.8-2.4T-A95B, the first Qwen-Max-class model released as open weights, has been deployed on Amazon SageMaker HyperPod with vLLM, a sign the serving stack is finally catching up to the weight releases. (Source)
• Desert Ant Labs launched as a European frontier lab betting on opinionated, fast models that run on device rather than on someone else’s cloud. (Source)
• The New Yorker’s Infinite Scroll column profiles Flock’s license-plate camera network as infrastructure for a closely surveilled world with no exit — the argument city councils keep hearing and keep losing to. (Source)
• An “Econ Scenario Explorer v1.0” collects competing models of what the economy looks like through the AI transition, published as a map of disagreement rather than a forecast. (Source)
• A first-person write-up of buying Google Ads to distribute malicious software is a reminder that ad review is an availability filter, not a safety one. (Source)
• A card-network explainer covers what Visa and Mastercard actually do — the authorization, clearing, and settlement layer most engineers only ever touch through Stripe. (Source)

One Line

Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.

— An Anthropic pretraining researcher announcing their resignation on X, posting as @hilbertspaess