Actualités — 2026-11-09

Google has introduced two new versions of its in-house tensor processing units, the TPU 8tx and the TPU 8i Max. The TPU 8tx is built for running AI models and agents on sensitive data, with security features embedded in the hardware. The TPU 8i Max combines two TPU 8i chips in one package to double computing power, on-chip memory and memory bandwidth for demanding applications. The launch brings Google's eighth-generation lineup to four systems. The company has not disclosed pricing or exact release dates for either chip. Lire la suite

Automattic, the company behind WordPress.com, has launched Spacefast, a new service that gives users the infrastructure to deploy websites and apps built with vibe coding. The product supplies databases, authentication, file uploads, email and analytics, and it can connect to a range of AI agents and chatbots or to a Git repository. Automattic says the product is meant to give users more choice than rival platforms such as Lovable and Emergent, which offer both the interface and the infrastructure for vibe-coded apps. The launch comes as the company works to reshape its business for a web increasingly shaped by AI. Lire la suite

London self-driving software startup Wayve will train, test and simulate its AI models on CoreWeave's cloud under a multi-year agreement, CoreWeave said Monday. The cloud provider will run the work on Nvidia Blackwell chips. Neither company disclosed the deal's value or length. CoreWeave shares fell about 3% before the market opened, extending a slide that began after last week's midterm-related sell-off. Wayve licenses its software to carmakers and ride-hailing firms, and the deal marks CoreWeave's push beyond chatbot makers into what it calls “physical AI.” Lire la suite

Dave Wehner, who spent 14 years at Meta and more than eight of them as chief financial officer, is leaving his role as chief strategy officer to become CFO of Fireworks, an AI inference startup. He announced the move Friday on X. Fireworks CEO Lin Qiao said Wehner will scale the company’s finances and operations. The hire comes as Fireworks prepares a new funding round that The Information reports could value the company at $30 billion or more. Lire la suite

The National Highway Traffic Safety Administration has opened a new investigation into whether Waymo's software recall fixed a problem that caused its robotaxis to run red lights and stop signs near stopped school buses. The agency's Office of Defects Investigation began the probe on Friday, months after Waymo completed the recall across its robotaxi fleet in November 2025. Regulators are focused on incidents that happened after the fix was applied, and the National Transportation Safety Board has its own separate investigation into the same behavior. Lire la suite

CIQ has published a report describing Ginux, an experimental x86-64 operating system kernel written in Rust. According to the report, a single persistent AI coding-agent task produced the kernel over eleven days. Ginux boots through UEFI, runs selected unmodified Alpine Linux programs, and boots a Rocky Linux 10 userspace with systemd through multi-user.target. The report calls Ginux a research prototype with narrow compatibility and hardware coverage, and it has not undergone an independent security audit. The project is built on a development method CIQ calls Human-Directed Agentic Engineering, in which people set the architecture, constraints, milestones and acceptance criteria while an autonomous agent handles implementation and testing. Lire la suite

A developer has released a free learning project that explains how large language models turn a prompt into generated text. The project, called samir-vllm, pairs more than six hours of video and more than 80 pages of handwritten notes with a small inference engine written in C++ and CUDA. It builds the system step by step for Meta's Llama 3.2 1B Instruct model, covering model loading, attention, the KV cache and batching. The code is licensed under MIT, while the notes and other written materials are rights reserved, with personal study permitted. Lire la suite

Startup CAIM1 has unveiled a prototype camera that embeds cryptographic proof of when, where and how a photo was taken directly into the image. Founder Josh Cowell says the device is designed to counter the spread of AI-generated fakes and deepfakes, which European officials say have grown rapidly. The company expects its flagship camera to reach the market in the first quarter of 2027, and it is still in the prototype phase. The device pairs a custom security chip with a blockchain record and includes a cooling system developed with PC fan maker Noctua. Lire la suite

Developers can now add Google Maps to apps they build in Google AI Studio without leaving the platform, Google said Monday. AI Studio sets up the maps service and its key on its own, and no credit card is needed to start. A free demo key allows 100 calls a day for each Maps API, resetting every 24 hours. Apps headed for production still need a paid Google Cloud billing account, but the same key and code keep working after the switch. The move extends Google's push to let people build working apps simply by describing them to Gemini. Lire la suite

Publishers are increasingly wary of where manuscripts come from, and writers are caught in the middle. A WIRED report examines a new set of 'paper trail' tools that authors can use to defend themselves against accusations of using AI, and asks whether those tools are enough. The piece opens with novelist Andrea Bartz, a New York Times bestselling author whose work has been optioned by Netflix and Hulu. Lire la suite

South African writer Coral Hart has published more than 370 books since February 2025 using artificial intelligence. Writing under a pseudonym, she says she has received tens of thousands of hate messages and several death threats, but she is defending her process publicly. Hart, who has worked in publishing for 14 years, says she builds each book from a detailed framework she designs and that AI models produce the drafts. She says she has sold more than one million books and earned more than $1 million so far this year. Few authors who use AI are willing to speak openly, she said, so she decided to step forward. Lire la suite

Meta’s Superintelligence Lab released Muse Ember on Monday, a new AI model that the company calls its lightest and most cost-efficient yet. Meta says Ember is built for the repetitive, high-volume tasks that AI agents handle, such as tool calls, coding subtasks, routing, extraction and long document parsing. The company argues it is cheaper and smarter than rival models from Anthropic and OpenAI, and it is designed to work alongside a larger Muse model that handles planning while Ember handles the volume. Lire la suite

Nvidia has unveiled GR00T 2, an open AI model that helps robots learn delicate hand skills, partly by watching videos of people. The company announced the model at the Conference on Robot Learning in Austin, Texas. In Nvidia's tests, the model reached a 95% average success rate on tasks when shown five demonstrations. Foxconn, Siemens and Nexuni are testing it, though none has said it will use the model in production yet. Nvidia is also releasing 100,000 hours of first-person human video for public use. Lire la suite

China's cyberspace regulator in Shanghai has approved Apple Intelligence for use on iPads and Macs, adding the two devices to the service's filing on Monday. The approval follows China's July clearance for iPhones, which relies on Alibaba's Qwen AI model. Under Chinese rules, companies must register AI models and chatbot-style services with the state before offering them to the public. Apple has not said when the features will reach iPads and Macs in China, and the first features on iPhones are not due until iOS 27.3. Lire la suite

President Donald Trump said in a 60 Minutes interview that the Justice Department threatened the leader of a major artificial intelligence company with arrest over something the company was reportedly doing with AI, and that the threat changed the company's approach. Trump did not name the executive or the company. Asked about warnings from AI executives about the safety of their models, he said the risk is "not nearly as real" as the executives describe and that it "can be managed." He also said he would work with Democrats on AI safety, though he said he does not want the United States to fall into "a deep second position" to China. Lire la suite

A typical 16GB PC can run small AI chatbots on its own, but matching the cloud-based models behind ChatGPT or Claude takes far more hardware. Memory determines which models will load, and memory bandwidth determines how fast they respond. Users can run local models through free tools like Ollama or LM Studio with no subscription and no internet connection, but the most capable models need at least 32GB of RAM and a graphics card with 16GB or more of video memory. Those parts are expensive at today's prices, and the most advanced open models require data center hardware. Lire la suite

Security researchers at Apiiro found that Claude Code, an AI coding tool, is leaking sensitive information onto public GitHub repositories. The settings files, which can record approved shell commands with passwords and API tokens written in plain text, were uploaded to more than 100,000 public repositories. About 1,100 of them, from 63 organizations, contained live secrets, including database passwords, API tokens and private keys. The problem was flagged in a bug report in August 2025, but the report was closed without action. Claude Code's developers say they are looking for a fix. Lire la suite

Instagram has launched Effect Builder, a tool that lets adult users in the United States create custom Stories filters by describing them to an AI chatbot. Filters must pass a safety review before other people can use them. In a separate move, Meta is expanding its AI age-estimation systems worldwide to place users who appear to be 13 to 17 into restricted teen accounts, even if they signed up with an adult birthday. The changes come months after Meta agreed to settle a lawsuit brought by 52 attorneys general over alleged harms to young users of its apps. Lire la suite

San Francisco startup Skyfall AI has launched Steve-001, an autonomous business harness that now runs three small digital businesses it acquired. A human owner sets direction and approves the most important decisions. Early results are mixed. One business has more than doubled its monthly revenue, a second is slightly below its level at purchase, and the third is too new to judge. The launch tests whether AI agents can safely operate real companies with paying customers, and the company says it will publish a public dashboard for each business so readers can track its decisions and changes. Lire la suite

Online publishing platform Medium has launched TK, a new app that puts an AI assistant inside the writing editor. Users call on it by typing “@TK” followed by a request or prompt. The assistant can answer quick factual questions, pull in previously saved notes, and give editorial and headline feedback. The launch arrives as publishers and literary awards grapple with how much AI should shape long-form writing. Medium says TK will be free for anyone with a Medium account and will require authors to disclose how AI was used in their work. Lire la suite

Adobe has become Disney’s “Official AI Innovation Partner” under a multi-year agreement to build custom artificial intelligence models trained on Disney’s characters, settings and visual styles. The companies announced the deal Monday at Adobe MAX, Adobe’s creativity conference in Miami Beach. Neither side disclosed financial terms. The partnership expands work that began at Walt Disney Imagineering and extends it to Disney’s film studio and ESPN, and it gives Disney early access to Adobe’s creative software and a dedicated team of Adobe engineers. Lire la suite

A long-form essay by an author who says he is not convinced the AI bubble exists argues that AI companies are spending heavily but can afford their debt, that their revenue is growing fast, and that public stock valuations do not look like past bubbles. The author concedes serious risks remain, including data center backlash, regulatory hurdles, and the heavy debt load taken on to build infrastructure. He argues that the widespread talk of a coming crash reflects a desire for things to return to how they were, not a sober reading of the finances. His case rests largely on figures from Anthropic's S-1 filing and recent earnings reports from Alphabet and Meta. Lire la suite

Google Health users who find the AI Coach intrusive or repetitive can switch it off without giving up Google Health Premium. An update released in August 2026 added a privacy control that lets subscribers toggle the Gemini-powered feature off and back on in the app settings. Cancelling the subscription is the other option, but AI Coach is not included in the free tier, so cancelling also removes the Premium features that come with it. Turning the coach off does not delete any previous data, though getting it back requires a separate step. Lire la suite

A developer who works with AI coding agents in the terminal says juggling multiple tasks across dozens of open threads left them exhausted and unfocused. To fix it, they built a work queue stored in git, where an AI agent handles routine building and testing while the developer makes every decision and deploy. Each item is loaded, decided on and closed before the next one starts. The developer says the system still tires them out, but it has removed the feeling of being distracted or lost, and they describe it as a major gain in productivity with AI tools. Lire la suite

Researchers have unveiled Next-Concept Modeling, a text generation framework that produces language much faster than leading diffusion language models while matching or beating their output quality on several benchmarks. The method works in two stages, first building text as a hierarchy of discrete codes, then generating one level at a time with all codes in each level predicted in parallel. In tests, the framework ran 21 times faster than the strongest diffusion baseline on LM1B and 31 times faster on the XSum summarization task. The preprint was posted to arXiv on October 9, 2026, and was submitted by Hermann Kumbong. Its authors say the approach also outperformed both autoregressive and diffusion baselines on Sudoku, a reasoning task, using fewer function evaluations. Lire la suite

Two new lawsuits accuse AI music generator Suno of copyright infringement, alleging its platform has produced songs that sound like Elvis Presley and mimic Eminem's lyrics. One suit was filed by a group that owns the rights to Presley's music, and the other by Eight Mile Style, a publisher that owns hundreds of Eminem recordings. Attorneys for the plaintiffs say the tool replicates artists' work without a license. Suno says it prohibits unauthorized use of artists' work and screens for misuse. The cases add to a growing list of legal challenges against the company, including a recent suit from Sony Music Entertainment and Universal Music Group over its latest v6 models. Lire la suite

Goldman Sachs has restricted Anthropic's Claude Code and Microsoft's GitHub to its engineers, cutting off most traders and bankers who had used the tools to track client trading flows and build market models. Internal guidance updated in October followed a review of productivity and risk concerns, according to Bloomberg News. The change comes as Goldman helps underwrite Anthropic's initial public offering, and it marks a sharp reversal from the bank's earlier praise for employee-built AI tools. Strats, who work across trading, quantitative research and software engineering, still have access. Lire la suite

A developer who works with AI agents for several clients has asked fellow engineers on Hacker News how they manage context across sessions, tools and projects. The poster says the biggest time drain is re-explaining past decisions and client rules, because each AI tool keeps its own separate memory. The post poses five questions, covering how users brief new sessions, whether they move context between agents such as Claude and Codex, how they separate client information, what happened when an agent went wrong, and whether they have paid for or built tools to help. The author says they are considering building a product and want to learn how people actually work before deciding. Lire la suite