1. The interviewer asked candidates to bring the parts they were building at Apple The hardware chief allegedly told Apple engineers interviewing at OpenAI to arrive with something specific: the components they were working on, and unreleased product samples.

2. Apple rebuilt the iPhone around Siri; Google rebuilt Waze around Gemini Apple's iOS 27 shipped its first public beta today, and the headline change isn't a new app.

3. Lorde Told a Festival Crowd AI Glasses Aren't Sexy. The Sponsor Made Ray-Ban Meta The pop musician Lorde paused her set at the Mad Cool Festival in Madrid on Thursday to attack AI glasses from the stage. She named no brand.

In Brief

• Anthropic localizes Claude pricing for India Anthropic rolled out rupee-denominated Claude subscriptions in India, its largest market after the US. The move drops the friction of dollar billing for individual and business users.

• Altman mocks Musk's orbital data center plans Altman fired back at Musk, who had called him a scammer, by ridiculing Musk for selling public-market investors on near-term space data centers. Most experts agree orbital compute remains impractical on Musk's timeline.

• Nadella warns enterprises against locking into proprietary models Nadella cautioned companies that depend on closed models sold by large AI labs. The worry among Silicon Valley builders is that proprietary model vendors function as Trojan horses inside customer stacks.

• Defenders weaponize prompt injection against attack agents Security teams now plant prompt injections to fight back. A technique called "context bombing" tricks autonomous hacking agents into shutting down before they can cause damage.

• MIT Technology Review checks Anthropic's latest research claims The publication broke down what Anthropic's newest AI finding actually demonstrates and where it falls short. Anthropic, valued near $1 trillion, is known for heady research, including studies on whether models can feel pain.

• Ars Technica maps what world models can and cannot do Researchers explained how world models simulate environments, their current uses, and the open questions. The piece separates working capability from unproven promise.

• TechCrunch probes the limits of fully user-aligned AI The piece asks what happens when an AI serves only its user's intent, using the extreme case of helping someone commit murder. It tests where user alignment collides with public harm.

• Long-Horizon-Terminal-Bench grades agents on partial progress Researchers released a benchmark of 46 long-horizon terminal tasks across nine categories. Dense reward grading scores intermediate steps, replacing pass-fail final-outcome checks that hide partial solutions.

• Paper pitches text-to-video pretraining as a route to general vision models The authors argue large-scale video generation supplies the spatiotemporal priors and vision-language alignment needed for general visual intelligence. They introduce a model built on that pretraining paradigm.

Read the full edition →