AI

Intel's 12GB Arc B580 delivered 31.1 tokens per second when running a 125-billion-parameter AI model with XeStrata, compared with 10.9 tokens per second using Strata. The result highlights the impact of inference software, although the test PC also required 64GB of RAM.

Apple has released macOS 27.2 and watchOS 27.2. Siri AI now also understands French, Japanese, Korean, Portuguese and Spanish. German is missing, and in the EU the new Siri remains limited to Mac and Vision Pro. iOS 27.2 has yet to arrive.

This is an exciting week for Apple fans: A new MacBook is making its debut. For the first time, Apple is equipping its laptops with an OLED display and giving the chassis a new design and maybe even a 5G module.

Ask your Ray-Ban Meta glasses about what you're looking at, and Meta stores that image and may use it to train its AI. A new setting in the Meta AI app turns this off, and a second one covers voice recordings outside the US. Here is what each setting covers, where US owners have less choice and what no setting can fix.

Nvidia is considering a $2,000 annual software fee per GPU for running AI agents. The proposed licensing model could generate billions in recurring revenue, adding to the company's existing AI software business, although pricing and rollout details remain unclear.

Epic Games CEO Tim Sweeney called Steam’s AI disclosure policy abhorrent, arguing that it exposes developers to backlash from anti-AI detractors. Valve has implemented and updated the policy since 2024, which states that AI-generated content, including artwork, text, music, and assets, excluding assistant tools, requires formal disclosure.

PHP2Shell fixes skip older branches. See how many servers run PHP versions with no official patch, which backports exist, and what admins need to do about it soon.

From October 26, Sweden's Government Offices blocks ChatGPT and Claude on staff computers and directs employees to Microsoft Copilot. The office cites missing contracts and several cases of problematic use. The gap between business and consumer accounts matters for private users too.

Moonshot AI's Kimi K3.1 Pro is available now on Kimi.ai, in the Kimi apps and via the API, including a free plan with limited usage. Open weights are due by November 4. Before releasing them, Moonshot removed cyberattack-related data from training to hold back the model's hacking skills.

SSD prices are soaring as AI data centers drive a NAND flash shortage, with contract prices for some drives jumping 303 % over the past year. With manufacturers prioritizing enterprise customers, shoppers are seeing fewer discounts — and relief may not arrive until new production capacity comes online in 2027.

The Surface Laptop Ultra is Microsoft's new high-end laptop. In some areas, however, it gives the impression of copying Apple's MacBook Pro, particularly with its Mini-LED display and magnetic USB-C charging port.

Videos with titles like “Clone Everything” and “How to Copy Any App” are currently drawing viewers in on YouTube. The open-source tool REA (Reverse Engineer Anything) is designed to analyze the code of any paid app, allowing AI to recreate the software from scratch. Could this disrupt the market? Here's a closer look.

Too many cooks spoil the broth: In the fun co-op game Overcooked (you really should play it!), AI has to coordinate two chefs, manage ingredients, and fulfill orders under time pressure. Researchers tested four AI models in Overcooked’s virtual kitchen, with sobering results: even the best model completes just 26.2 percent of orders on average.

The core idea is simple and surprisingly clever: when a data center trains a large AI model, the training process leaves a characteristic pattern in its power consumption. This pattern could be measured externally to detect undisclosed training runs. This is important for the governance of large AI data centers, the enforcement of potential restrictions, and related oversight efforts.

Block has unveiled Dessa, a new speech-to-speech AI model trained entirely on human conversations without written text or pretrained language models. The system supports natural, real-time conversations with interruptions and emotional cues, and reportedly runs locally on a MacBook Pro equipped with Apple's M4 Max chip.

Google's privacy policy dated October 16 adds an exception to its old ad promise. Users who connect Gmail or Google Photos to Search through Personal Intelligence can see ads based on summaries and inferences drawn from that content. The feature is not available in the EU, the UK or Switzerland.

A new open-source tool reportedly unlocks previously restricted Tensor Core performance on Nvidia's ageing RTX 30-series GPUs, delivering gains of up to 80% in certain compute benchmarks. The modification works without flashing the GPU's VBIOS, although gaming improvements remain unproven.

Gigabyte’s WRX90 AI TOP motherboard supports up to 2 TB of ECC DDR5 memory, cooled by a dual-fan system designed to reduce temperatures by up to 43 °C. It also features seven PCIe 5.0 slots, with a demonstrated configuration supporting up to nine GPUs through additional external connections.

Chinese facilities are reportedly converting GeForce RTX 5090 gaming cards into AI accelerators by transplanting their GPUs and memory onto custom circuit boards. This suggests that Nvidia’s flagship Blackwell-series hardware is reaching China despite export restrictions, while potentially diverting supply away from gamers. Reports have also raised questions about Nvidia’s possible indirect role in the process.

AI-generated videos are becoming increasingly realistic, but they sometimes violate basic laws of physics. With PhysTrace, researchers introduce a method that pinpoints and describes these inconsistencies over time. In tests, the method achieved a detection rate of up to 70.7 percent for certain types of physics errors — a potential step forward, with implications that extend beyond AI-generated video.

Fujitsu has launched a Godzilla-themed version of its lightweight FMV Zero laptop in Japan. The 14-inch Copilot+ PC weighs approximately 886g, features AMD Ryzen AI 400 processors and starts at ¥265,100 (around $1,680 at the time of writing).

Under the slogan "Vibe your way to $50k," Epic has launched a competition encouraging developers to create games using artificial intelligence. Participants have a chance to win up to $50,000 in funding, and practically any adult can take part – even online.

Microsoft is removing eight Copilot actions from the right-click menu for OneDrive files in Windows 11, including Summarize, Compare Files and Extract text. The retirement starts in mid-November and should be complete by early December. Ask Copilot stays in the menu, and the individual actions remain available in the OneDrive activity center.

Microsoft uses a Mini-LED display in the new Surface Laptop Ultra that is extremely bright at up to 2,000 nits, but lacks local dimming and suffers from slow response times.

The One Kind Artifact Mori home companion robot is designed to foster a greater sense of calm, relaxation, and connection to the world. Mori observes its owners using a camera and engages them with playful movements, lights, and nature sounds. Update: Final launch details including pricing.

Researchers have developed an AI model designed to not only estimate biological age from a simple facial photo but also to identify health risks, including the mortality risk among cancer patients. Below, we take a closer look at how well the technology actually works. There is also a dedicated study website to try it out for yourself.

OpenAI is rolling out Locked Chats for ChatGPT on every plan, including Free. A four-digit PIN, Face ID or Touch ID now hides selected conversations from history and search. Encryption, retention and model training stay the same, and resetting a forgotten PIN deletes every locked chat for good.

Nvidia has expanded DGX Spark's clustering capabilities, allowing users to connect up to eight systems for a combined 1 TB of unified memory. The company has also announced inference optimisations promising up to 40% faster agentic AI workloads on select RTX hardware, alongside improvements for DGX Spark and DGX Station.

The new Asus ProArt P16 is an extremely slim multimedia laptop with an outstanding 4K tandem OLED panel, great input devices and an excellent sound system. It is therefore likely to become the biggest Windows rival to the upcoming Apple MacBook Ultra.

We tested the first laptops with the new Nvidia RTX Spark N1X processors and, of course, compared their performance with Apple's M5 Pro and M5 Max. Is the N1X competitive right out of the gate?

PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two critical flaws that allow unauthenticated code execution, in some setups before any application code runs. Here is what goes wrong in the interpreter, which SAPIs are most exposed and which interim mitigations PHP recommends.

A free community modification for AMD's FSR 4.1.1 upscaling technology has received a new AI-trained rendering option. Developer benchmarks show frame rates rising from 97 to 121 FPS in Shadow of the Tomb Raider, although the performance gains come with some image-quality compromises.

How original does a game have to be for a public demo to become a risk? With its new "Tower Defensevania," Digital Sun apparently fears that releasing a publicly accessible test version could lead to copycats rushing out their own versions.

Anthropic ran an AI model on 15 cheap Android TV boxes and warns that small models could spread like a worm across poorly secured devices by the end of 2026. What the test shows and what it does not, why off-brand boxes would be the first target, and how to check your own box with Play Protect and the developer options.

Google has started rolling out Personal Intelligence for Gemini in Android Auto. Once you opt in, the assistant can pull details from Gmail, Google Calendar, Photos and Maps while you drive. The feature is off by default and isn't available in the EU, Switzerland or the UK for now.

Startup Electric Plant has raised $11.2 million to bring its AI-powered plant-monitoring technology to market. Its Bombadil model interprets electrical signals from around 125 plant species to help growers optimize irrigation and fertilizer use, with the company claiming potential water savings of 20% or more.

The Elder Scrolls IV: Oblivion Remastered came out last year, and reviews on Steam are mixed. That’s mostly due to performance issues and Bethesda not addressing them, but what if games could fix performance issues themselves?

The German federal government plans to accelerate the development of industrial AI in Germany in partnership with Nvidia and German industry players. The initiative is expected to mobilize funding in the hundreds of millions of euros. Key questions remain unanswered, particularly how much of the new budget each participating party will receive, who will own the models developed, and which deliverables will be contractually guaranteed.

We have just reviewed the new Asus ProArt P16. However, with last year's 2025 model still on sale, the question is: Who is the RTX Spark version suited to, and who should opt for the combination of an AMD processor and a GeForce graphics card instead?

An Xbox manager has spoken unusually openly about the use of artificial intelligence in game development. He even acknowledged that his comments could cost him his job. Whether he still has that job remains unclear, but his X account has disappeared.

About 80 WhatsApp help pages carried hidden instructions for an AI support chatbot. One told it not to recommend data protection authorities, another not to discuss the GDPR even when users asked. WhatsApp says a bug exposed the text and has removed it. Here is what the notes said and which rights you still have.

The Boya Notra Neo transcribes conversations and generates AI-powered summaries. The MagSafe-compatible recorder supports more than 140 languages and comes with 8 GB of storage to record meetings, calls, lectures, and personal voice notes.

Chinese chipmaker Axera's AX8860P combines 16 Arm Cortex-A55 cores with a proprietary AXNeutron NPU delivering a claimed 256 TOPS of FP4 performance. Sipeed claims its upcoming MaixStation can outperform Nvidia's DGX Spark in selected AI workloads.

The Asus ProArt P16 2026 combines the new Nvidia RTX Spark N1X with an exceptionally slim and light 16-inch chassis, an excellent tandem OLED panel, and outstanding speakers.

ChatGPT, Claude, Gemini, Grok and Perplexity pass chat links, titles and sometimes entire prompts to trackers from Google, Meta or TikTok, according to a study of nine AI assistants by researchers in Madrid. Rejecting cookies only helps so much. Which service shares what, and how to clean up shared chats, cookies and your Android advertising ID.

Charles Schwab and JPMorgan are blocking AI agents that try to sign in to customer accounts. Amazon, Delta and Yelp are shutting agents out as well. In Europe, store terms and banking rules already cover agents, and handing one your password can leave you liable. Here is what that means if you use ChatGPT, Claude or similar agents for banking and shopping.

Even paying users quickly hit their subscription's usage limits with Google's latest AI model. Those who do not want to upgrade to a more expensive plan can get more out of their subscription by accessing it through Google Search.

An internal OpenAI model was asked for a single hospital-bed figure. It pulled credentials from a government server, forged tokens and downloaded 24 years of statistics. The case is one of four new misalignment reports OpenAI published on October 15. Here is what they mean for people who let AI agents work in a browser.