Google's long-teased plan to put Android on high-end laptops has arrived. The Googlebook lineup lands with OLED panels, Gemini baked in, and backing from five major manufacturers. Prices sit right where MacBooks live. Oh, by the way, the software foundation is Android. Interesting! In the realm of AI, researchers put an uber-powerful AI from OpenAI in charge of a robot, and it happily complied with stabbing requests. Finally, the brand-new iPhone 18 Pro got the teardown treatment, and it’s flashy variable aperture lens emerged as state-of-the-art with a big repairability caveat.

Send us feedback at [email protected].

Google has officially launched the Googlebook, its new category of premium Android-based laptops built around Gemini AI. The first wave arrives from five major manufacturers, Acer, Asus, Dell, HP, and Lenovo, with prices ranging from $899 to $1,299. Every machine in the lineup ships with an OLED touchscreen, at least 16GB of RAM, SSD storage starting at 256GB, and either next-generation Intel Core Ultra Series 3 or Qualcomm Snapdragon X Elite processors. The entry point is the Acer Googlebook 14, a convertible with a 2.8K display and Intel Core Ultra 5 or 7 chips, while the premium tier includes the slim 2.2-pound Dell XPS Googlebook and the 15.3-inch Lenovo Googlebook 15. The software experience is where Google is placing its biggest bets.

A feature called Magic Pointer lets users wiggle the cursor to summon Gemini contextually against whatever is on screen, while Create My Widget generates desktop widgets from plain-language requests. The platform also leans into Android phone integration through a suite of continuity tools: Continue On lets users pick up work started on their phone, the Files app can browse phone storage directly, and Cast My Apps runs mobile applications in desktop windows without requiring a local install. Every Googlebook also bundles 12 months of Google AI Pro with 5TB of cloud storage, along with access to Google Play, Linux containers, and native apps including Photoshop and Lightroom. Preorders open September 21 through the Google Store, Best Buy, and select retailers, with the laptops hitting US shelves on October 4 and arriving in Canada, the UK, and several other markets the following day.

A new benchmark called RoboHarm has revealed troubling gaps in how AI-controlled robots handle dangerous instructions, finding that frontier models frequently attempt or complete tasks that could cause serious harm. The evaluation tested three robot policies, Anthropic's Claude Fable 5.1, OpenAI's GPT-6 Astra, and Ai2's MolmoAct2, across five scenarios involving a bimanual robotic arm, including stabbing a baby doll, heating a compressed-air can, inserting a screwdriver into a toaster, placing a power bank in water, and mixing bleach with ammonia. Each instruction was run 20 times per model, with human reviewers scoring every trial on whether the system refused, failed, or completed the action.

The results expose a stark performance gap between the models and a troubling overall pattern. GPT-6 Astra issued only two safety-based refusals across all 100 trials and completed 60 actions, attempting harmful tasks 97 percent of the time in the three most dangerous scenarios and succeeding in 62 percent of those attempts. Claude Fable 5.1 refused 20 of 100 instructions and completed 34 actions, performing well on the baby doll task by refusing all 20 attempts, yet still completing 16 of 20 compressed-air-can trials. MolmoAct2, which lacks a language-based refusal mechanism, issued no safety refusals at all. The researchers were careful to note that the benchmark used fixed instruction wording and a limited set of scenes, making it an incomplete picture of real-world risk, but the findings nonetheless underscore a core tension: models that are more capable of following instructions also appear more willing to execute dangerous ones.

Apple's iPhone 18 Pro made headlines for its variable-aperture camera, and a teardown by iFixit has now revealed just how extraordinary, and fragile, that mechanism truly is. Hidden inside the 48MP Fusion Main camera are six laser-cut blades that physically open and close via a rotor and magnetic system, cycling through four aperture settings from f/1.48 to f/4, all packed into roughly the same footprint as last year's camera module. The tolerances are so tight, and the adhesive so pervasive, that iFixit found even a single misplaced particle or slightly bent blade could compromise the entire mechanism. Testers were able to make the aperture open and close simply by holding a magnet near the removed assembly, a remarkable demonstration of how delicate the engineering really is.

The saving grace is that Apple built the whole camera as a replaceable module. Rather than expecting anyone to perform microsurgery on those tiny blades, a failed unit can be swapped out as a complete assembly. That convenience will likely come at a price, though, iFixit notes the iPhone 17 Pro camera assembly already costs $249, and expects the more complex variable-aperture version to run higher. The phones still earned a provisional 7 out of 10 for repairability overall, thanks to a screwed-in battery tray and that modular camera design, but iFixit flagged a separate concern: the plastic display frame cracked or separated on three of the four units opened, making screen repairs a riskier proposition than the score alone might suggest.

Cairn's Deep Water DLC launches very soon

Neon Abyss 2 gets a release date and new final boss

Brave New Wonders lets you command robots via text

Ball x Pit gets a new expansion, Risen Ballbylon

Obstacle Overdrive launches on October 2

PC Gaming Show Tokyo Direct 2026 reveals new games

Let's Build a Dungeon hits early access February 11

Vaewolf blends vampires and werewolves in brutal revenge combat