Hardware
iPhone Duo preorders slip into 2027
Delivery estimates for Apple's first foldable pushed to late January within an hour of preorders opening, though in-store stock may arrive sooner.
Still
Crunchy.
YESTERDAY, STILL WORTH KNOWING
Friday brought a foldable iPhone nobody can get yet, an interpretability row splitting AI researchers, and a fresh push to make game purchases mean something.
The iPhone Duo sells out into next year
Apple opened preorders for the iPhone Duo, its first foldable, at 8AM ET on Friday, and within an hour every configuration on the website was showing late January delivery dates. That means many online orders will not arrive until 2027. The phone still reaches stores on October 23rd, and Apple typically holds back stock for walk-in buyers, so visiting an Apple Store may be the quicker route for anyone determined to get one early. Whether the delay reflects high demand, low supply or both is unclear.
The Duo uses the shorter passport-style design Samsung introduced earlier this year and starts at $1,999 for the 256GB model in star white or night sky, a little above the Pixel 11 Pro Fold and Galaxy Z Fold 8.
An interpretability researcher rewrites his own theory
Nick Cammarata of Goodfire, who helped shape the field's dominant theory, has published an essay arguing that large AI models are far easier to understand than researchers assumed. He now rejects superposition, the idea that networks cram unrelated concepts into the same neurons when short of space, and reinterprets the famous cat-and-car neuron as a V-shape detector instead. As evidence, he reports that roughly 90 percent of neuron poles in Qwen3-32B have one main job, with only 2 to 4 percent unexplained, and much of the analysis was done by AI agents working in bulk.
Reactions split sharply: OpenAI's Roon called it potentially a major moment in scientific history, while DeepMind's Neel Nanda warned against overhyping it, and Roon himself noted the results need independent confirmation. The stakes go beyond curiosity, as Cammarata argues interpretability is a prerequisite for alignment, letting humans spot deception directly in a model's circuits.
Open models: China leads, the West chases, Ubuntu ships one
Reflection launched Beam, its first model, as open weights under Apache 2.0: a 501B-parameter sparse mixture-of-experts model with about 23B active per token, a 1M-token context window and a coding and agentic focus. It needs the memory of a 501B model but computes at roughly 23B speed, and Hugging Face is offering day-zero support plus $10,000 in Hub credits for the best projects. Yet experts at a Rest of World event said Western open models, including Beam, still trail China's best on performance, with four Chinese models in OpenRouter's top five and the leading Chinese systems just 3 percent behind US rivals on LiveBench. Cost, local fit and trust all matter: open models help users sidestep wariness of both Huawei-style scrutiny and American access restrictions, and Xi Jinping has made open models central to China's 2030 AI ambitions.
Meanwhile Ubuntu 26.10, codenamed Stonking Stingray, arrived with Myna, an opt-in speech-to-text agent that runs entirely on-device, plus GNOME 51, kernel 7.3 with Nova Lake and Zen 6 support, and post-quantum package signing. Canonical calls it Ubuntu's first AI integration, with accessibility the priority.
Gaming's ownership fight, and Linux keeps winning
New York representative Ritchie Torres introduced the Power To The Players Act, which would have the FTC treat game purchases as ownership rather than revocable licences and require an offline mode before servers shut down, citing The Crew and NBA 2K25. Sega, meanwhile, pulled the Total War: Shogun 2 Complete Edition remaster after 48 hours amid an Overwhelmingly Negative Steam rating, with complaints covering bugs, performance and artwork users believe was mangled by AI upscaling; the original returns to sale and existing owners keep access.
On Linux, Epic confirmed it is working on Fortnite, but only via the Epic Games Store launcher if enough distro maintainers meet security requirements designed to bring Easy Anti-Cheat closer to Windows-level protection. Valve shipped a 64-bit Linux Steam client in beta alongside achievement sets, and Linus Torvalds posted his first LLM-assisted kernel patch, claiming a 23x gain in a synthetic benchmark on an old cache contention problem while cautioning real-world impact will be far smaller. Leaked die shots of Intel's Nova Lake-HX and H mobile chips rounded out the hardware news, showing a six-die design mixing Intel 3-E, Intel 18A-P and TSMC N2P.