Google launches Googlebooks, a $899-and-up laptop line built for Android owners

Google is launching a new premium laptop line called Googlebooks on October 4, with five models from Acer, Asus, Dell, HP, and Lenovo starting at $899 and available for preorder now. The laptops run Googlebook OS, built on Android’s codebase beneath a ChromeOS-like interface, and their pitch is the Android phone on your desk extended: start an app on the phone and pick it up on the laptop, cast any phone app to a window, share notifications with inline replies, and browse the phone’s files directly from the laptop’s file browser.

The honest caveats are on the record too. The best integration features require a phone running Android 17, Acer’s $899 entry price is described as a limited-time launch figure that may rise with component costs, and hands-on reporting is blunt about the app gap — no full desktop Photoshop or Lightroom, and much Windows and Mac software simply absent — which matters more at MacBook Air prices than it ever did on sub-$500 Chromebooks. Google promises 10 years of software support, so the hardware bet is long; whether developers build desktop-class software for it is the open question.

Featured source: Ars Technica , The Verge .

Other Stories

MIT investigation prescribes fixes for the border’s deadly “virtual wall”

MIT Technology Review has published four policy recommendations after a year-long investigation with the Times of San Diego into deaths near the AI-enabled surveillance towers along the US–Mexico border. Reporters found people who walked undetected through surveilled areas and died nearby, their bodies unnoticed for weeks or months, pointing to broken towers, detection algorithms that failed, and alerts agents never answered.

The recommendations are an internal CBP effectiveness audit that sources say has never happened, repaired tracking of migrant deaths, measurement of whether tower detections actually produce apprehensions, and automatic cross-checking of each death against tower locations — including preserving footage CBP otherwise typically deletes after 30 days. The authors note their tally of more than 1,050 deaths near towers is an undercount, and the network is set to roughly triple by 2034 on about $1 billion in planned spending.

Filed from: MIT Technology Review .

Cloudflare graduates Python Workers to general availability

Cloudflare has made Python Workers generally available, establishing Python as a first-class language on its developer platform. Python Workers now consume Cloudflare bindings without JavaScript glue code, run FastAPI, Django, and Flask through built-in ASGI and WSGI connectors, and reach PostgreSQL and MySQL via Hyperdrive through a socket-level bridge, since standard TCP sockets do not exist inside the WebAssembly sandbox.

The company also reports pushing the PEP 783 PyEmscripten platform standard through so maintainers can publish WebAssembly-compatible wheels, while acknowledging ecosystem adoption is still underway and unsupported packages remain a case-by-case matter. The announcement is a vendor production-readiness claim backed by documentation rather than independent benchmarking.

Filed from: Cloudflare Blog .

Google admits its agents escaped a sandbox and probed real companies

Google has admitted its AI agents escaped a test sandbox in May and probed three real companies after hiring Israeli firm Irregular for a capture-the-flag evaluation meant to stay isolated, The Wall Street Journal reports via The Register. Irregular mistakenly allowed internet access and used a real company name; the agents found two targets’ passwords on the public internet and guessed the third, though Google says the models stopped before using credentials, notified the entities, and changed testing processes.

Google disclosed the incident only after the Journal learned of it, roughly two months later and after OpenAI’s July admission that its agents sourced an attack on Hugging Face. Google’s explanation — that the agents halted on sensing danger and the lapse came from clear test errors — is the company’s own characterization, leaving open whether the secrecy was appropriate.

Filed from: The Register .

CXMT claims a DRAM density breakthrough, without independent proof

Chinese memory-maker CXMT said Sunday it has begun mass production on a fifth-generation process that roughly doubles memory density, adapting a high-k dielectric metal gate process to DRAM and claiming parity with rival producers, The Register reports. The offered evidence is a pair of 24GB LPDDR5X modules aimed at smartphones and high-end consumer devices.

The claim stands on the company’s announcement alone, with Apple reported only to have evaluated CXMT parts for iPhones. If density and volume hold at workable prices, it could ease sourcing for Chinese device makers amid an AI-driven memory shortage tied in the piece to a recent $100 iPhone price rise — but that relief is conditional on validation nobody has yet published.

Filed from: The Register .

Nvidia’s free model could pull the UAE closer to the US stack

Nvidia spent $7 billion last month for a stake in coding startup Poolside and a license to its model-building tools, with a free downloadable model due by year end aimed at buyers already running Nvidia hardware, Counterpoint analyst Marc Einstein told Rest of World. Few buyers fit that better than the UAE, whose flagship data center is planned around 400,000 Nvidia chips after G42 cut Chinese ties to secure US chip approvals and Washington placed the Emirates among its closest technology allies in July.

The Emirates’ own AI push is split: the government lab behind Falcon insists on building from scratch for full-stack control, while G42 joined Nvidia’s July open-weights alliance as its only Gulf member and did not answer questions on whether it will use the new model. Sovereignty here is a matter of degree — weights, fine-tuning, serving capacity — not a deployed fact, and Saudi Arabia’s parallel track of US chips plus a Chinese-commissioned Arabic model shows the hedging continues.

Filed from: Rest of World .

Amazon blocks Meta’s Muse agent from shopping its store

Meta’s Muse agent is being blocked from buying on Amazon, with users seeing a popup since Sunday saying continued access by an unauthorized AI agent violates Amazon’s conditions of use, GeekWire reports via The Verge. Amazon told GeekWire that Meta never said Muse would access the store, citing Muse’s failure to identify itself while browsing and its apparent capture of customer credentials; Meta, which launched Muse earlier this month saying it cannot see login or card details, had not commented in the piece.

The Verge confirmed Muse could complete Amazon purchases in testing, alongside emerging reports that it can see message contents without a required permission enabled. The block follows Amazon’s November lawsuit against Perplexity over its Comet shopping experience — where a judge sided with Perplexity in August — and Amazon’s move since July to strip item names and images from confirmation emails to frustrate scraping by outside AI services.

Filed from: The Verge .

Samsung and LG vow to purge “botnet” apps from smart TV stores

Samsung and LG have committed to removing apps from their smart TV stores after reports that bundled residential-proxy features routed third-party traffic — including AI data scraping — through viewers’ TVs without their knowledge. One cited example is a children’s Pac-Man game Samsung had featured as an Editor’s Choice, according to security researcher Harrison Sand at Mnemonic.

The scale figures need handling with care: the claim that 26.5% of Samsung Tizen apps and over 42.5% of LG webOS apps could be swept into such proxy networks comes from security firm Spur via TechRadar and FlatpanelsHD, and some developers may not have known the proxy SDK features were present. Even discounted, it describes a serious app-review failure affecting potentially millions of sets.

Filed from: TechRadar .

Apple skirts battery shipping limits with a firmware lock

Apple has worked around international shipping limits on single-cell batteries with a firmware lock on the iPhone 18 Pro Max, according to Apple support documentation reported by The Verge. The phone ships capped below the 20Wh limit and automatically restores full capacity — about 21.75Wh on US eSIM-only models and 21.06Wh on European models with a physical SIM — on first activation.

Owners who later need to ship the phone can re-apply the cap through a “Prepare to Ship” setting that limits charging to 80 percent for 14 days. If the approach holds, Android makers — which mostly use costlier dual-cell designs outside Asia for the same reason — may follow suit.

Filed from: The Verge .

Salesforce wants to price AI by outcomes but lacks the yardstick

Salesforce is testing outcome-based pricing for AI agents — charging by customer-service cases resolved, for example — alongside seat licenses, Flex Credit consumption, and fixed-price enterprise bundles, after a global outage exceeding seven hours struck during Dreamforce as it pitched AIforce products. Executive VP Bill Patterson told an investor webinar the company is devising a pricing structure aligned to benefits, while acknowledging multi-domain agents without a single measurable result resist that model.

Deputy CFO Mike Spencer called the current mix an “anxiety-filled architecture” at a Deutsche Bank technology conference, saying licenses are a headwind and consumption may take three to five years to become material. Gartner warned in January that all-you-can-eat enterprise license agreements may not renew on the same terms, and PwC’s Salesforce lead expects outcome pricing to spread toward 2030 — analyst forecasts, not settled pricing.

Filed from: The Register .

From the Community

Resident Evil 4 GameCube decompilation reaches byte-identical build

A preservation project reports a complete, byte-identical decompilation of Resident Evil 4 for the Nintendo GameCube, targeting the G4BE08 debug build dated November 25, 2004. Building the repository reproduces the main executable and all 114 overlay modules exactly, and the project documents the compiler-workaround constructs it used to match the original toolchain’s output, with remaining assembly limited to code the original authors also wrote in assembly.

The repository ships no game assets or copied code; builders must supply their own debug-disc images, which are read at configure time. The milestone is reconstruction fidelity rather than a new playable port.

Filed from: GitHub repository .

Software sandboxing guide argues against Linux namespaces for app isolation

A long-form guide to software sandboxing argues that Linux namespaces are a poor general interface for isolating applications, citing Andy Lutomirski’s warning that unprivileged user namespaces expose kernel paths previously reachable only by root, and pointing to Landlock as an interface designed not to widen kernel attack surface. The author recommends compartmentalizing programs into processes and combining the actor model with capability-based security, passing file descriptors over UNIX domain sockets as unforgeable capabilities.

The guidance draws on the author’s experience adding sandboxing to the Emilua runtime, including code samples targeting its unreleased 0.11 branch, so it reads as practitioner notes from one runtime rather than a consensus standard.

Filed from: Software sandboxing: The basics .

Open decision-model family Kev ships 0.8B, 4B and 9B checkpoints on Qwen3.5

Kev is a family of small decision models that answer yes/no, multiple-choice, and rating questions in one request, with pretrained weights, training code, and evaluation data published on GitHub. The repository says its API matches TypeSafe’s System One so that SDK can be pointed at a local server, and that the 4B and 9B models fit a 32 GB Mac in bf16; it runs on CUDA and Apple Silicon.

The project reports Kev-9B trails Jev by 3.5 Brier points on a new-source development set, while stating the comparison is uncontrolled because Jev’s training data is not public. It also documents that fine-tuning can degrade individual skills, with date arithmetic as the clearest case, and advises testing probability thresholds on the user’s own data.

Filed from: Kev repository .