Veni, vidi, ViSuisse

Les Écoles polytechniques fédérales de Zurich et de Lausanne ont travaillé pendant un an et demi pour sortir Apertus, un modèle qui se veut « 100% suisse », ouvert, respectueux du droit d’auteurs. L’idée est aussi d’avoir un modèle plus multilingue que ceux de la concurrence, en l’entrainant sur 40 % de textes non anglophones.

Apertus. Si comme moi, vous n’avez pas fait de latin, ça veut dire « ouvert » (comme dans apéro). Ici, les chercheurs suisses des Écoles polytechniques fédérales de Zurich et de Lausanne ont voulu afficher leur différence avec les modèles qui dominent le marché. Car, si la définition de l’IA ouverte divise encore le milieu, Apertus semble l’être particulièrement.

Un modèle qui se veut ouvert et respectueux des auteurs

Son objectif, selon nos confrères suisses du Temps : « offrir une alternative suisse, souveraine, aux modèles d’IA américains, chinois, voire européens ». Mais dans leur « rapport technique » [PDF], les chercheurs qui ont travaillé sur Apertus mettent aussi en avant le respect des auteurs et la volonté de proposer un modèle plus multilingue que les autres.

« Nous ne pouvons pas accepter que les bases de l’IA, si puissante, soient entre les mains d’une faction de multinationales », explique Antoine Bosselut, directeur du Laboratoire de traitement du langage naturel à l’EPFL, au journal Le Temps. En décembre 2023, les deux écoles polytechniques suisses ont donc créé la Swiss AI Initiative en s’appuyant aussi sur le Swiss National Supercomputing Centre. Celui-ci héberge le supercalculateur Alps composé, entre autres, de plus de 10 000 GH200 de NVIDIA.

Apertus est disponible en deux versions (8 milliards et 70 milliards de paramètres) sur Hugging Face sous licence Apache 2.0. Les chercheurs précisent que leur corpus de textes de pré-entraînement comme de post-entrainement « a été compilé uniquement à partir de données web, en respectant le fichier robots.txt non seulement au moment de l’exploration (janvier 2025), mais aussi en appliquant rétroactivement les préférences d’exclusion de janvier 2025 aux données web extraites lors d’explorations précédentes ». L’idée est d’être réellement conforme à l’AI Act européen et aux réglementations similaires, expliquent-ils.

Développer encore un peu plus le multilinguisme

Apertus a aussi été créé en utilisant, pendant l’entrainement, un mécanisme qui bloque la régurgitation des verbatims des textes utilisés pendant l’entrainement. Ainsi, les chercheurs évitent les éventuelles accusations de plagiat. Les chercheurs suisses ont voulu aussi se démarquer des modèles dominants en poussant autant que possible leur modèle vers un maximum de langues. Ainsi, leurs données d’entrainement contiennent des textes de 1 811 langues et 40 % ne sont pas anglophones. S’ils saluent les effort de leurs confrères qui ont créé les modèles ouverts BLOOM (sorti en 2022), Aya ou Qwen3, ils soulignent qu’Apertus a été entrainé sur 10 fois plus de langues que ces trois modèles.

• BLOOM : l’ambitieux modèle de langage de l’open science

Apertus ne joue pas au jeu du plus gros modèle possible. Rappelons qu’un modèle comme DeepSeek V3 a 685 milliards de paramètres. Mais, depuis quelque temps, on sait que cette course au volume commence à montrer ses limites. Apertus n’est pas multimodale, non plus. « Nous jouons dans une autre ligue, avec un modèle qui possède d’énormes qualités en termes de transparence, d’ouverture et de fiabilité. Et nous ne cesserons d’améliorer Apertus », assure Antoine Bosselut à nos confrères du Temps.

Pour les fonctionnalités de chat, le modèle peut être testé ici, sur une interface mise en place par le projet à but non lucratif Public AI Inference Utility.

Veni, vidi, ViSuisse

Les Écoles polytechniques fédérales de Zurich et de Lausanne ont travaillé pendant un an et demi pour sortir Apertus, un modèle qui se veut « 100% suisse », ouvert, respectueux du droit d’auteurs. L’idée est aussi d’avoir un modèle plus multilingue que ceux de la concurrence, en l’entrainant sur 40 % de textes non anglophones.

Apertus. Si comme moi, vous n’avez pas fait de latin, ça veut dire « ouvert » (comme dans apéro). Ici, les chercheurs suisses des Écoles polytechniques fédérales de Zurich et de Lausanne ont voulu afficher leur différence avec les modèles qui dominent le marché. Car, si la définition de l’IA ouverte divise encore le milieu, Apertus semble l’être particulièrement.

Un modèle qui se veut ouvert et respectueux des auteurs

Son objectif, selon nos confrères suisses du Temps : « offrir une alternative suisse, souveraine, aux modèles d’IA américains, chinois, voire européens ». Mais dans leur « rapport technique » [PDF], les chercheurs qui ont travaillé sur Apertus mettent aussi en avant le respect des auteurs et la volonté de proposer un modèle plus multilingue que les autres.

« Nous ne pouvons pas accepter que les bases de l’IA, si puissante, soient entre les mains d’une faction de multinationales », explique Antoine Bosselut, directeur du Laboratoire de traitement du langage naturel à l’EPFL, au journal Le Temps. En décembre 2023, les deux écoles polytechniques suisses ont donc créé la Swiss AI Initiative en s’appuyant aussi sur le Swiss National Supercomputing Centre. Celui-ci héberge le supercalculateur Alps composé, entre autres, de plus de 10 000 GH200 de NVIDIA.

Apertus est disponible en deux versions (8 milliards et 70 milliards de paramètres) sur Hugging Face sous licence Apache 2.0. Les chercheurs précisent que leur corpus de textes de pré-entraînement comme de post-entrainement « a été compilé uniquement à partir de données web, en respectant le fichier robots.txt non seulement au moment de l’exploration (janvier 2025), mais aussi en appliquant rétroactivement les préférences d’exclusion de janvier 2025 aux données web extraites lors d’explorations précédentes ». L’idée est d’être réellement conforme à l’AI Act européen et aux réglementations similaires, expliquent-ils.

Développer encore un peu plus le multilinguisme

Apertus a aussi été créé en utilisant, pendant l’entrainement, un mécanisme qui bloque la régurgitation des verbatims des textes utilisés pendant l’entrainement. Ainsi, les chercheurs évitent les éventuelles accusations de plagiat. Les chercheurs suisses ont voulu aussi se démarquer des modèles dominants en poussant autant que possible leur modèle vers un maximum de langues. Ainsi, leurs données d’entrainement contiennent des textes de 1 811 langues et 40 % ne sont pas anglophones. S’ils saluent les effort de leurs confrères qui ont créé les modèles ouverts BLOOM (sorti en 2022), Aya ou Qwen3, ils soulignent qu’Apertus a été entrainé sur 10 fois plus de langues que ces trois modèles.

• BLOOM : l’ambitieux modèle de langage de l’open science

Apertus ne joue pas au jeu du plus gros modèle possible. Rappelons qu’un modèle comme DeepSeek V3 a 685 milliards de paramètres. Mais, depuis quelque temps, on sait que cette course au volume commence à montrer ses limites. Apertus n’est pas multimodale, non plus. « Nous jouons dans une autre ligue, avec un modèle qui possède d’énormes qualités en termes de transparence, d’ouverture et de fiabilité. Et nous ne cesserons d’améliorer Apertus », assure Antoine Bosselut à nos confrères du Temps.

Pour les fonctionnalités de chat, le modèle peut être testé ici, sur une interface mise en place par le projet à but non lucratif Public AI Inference Utility.

Chaque sortie Noctua est un petit évènement, pour deux raisons : la première est le retour avec plaisir du beige et du marron, et la seconde est que le produit a souvent plus ou moins du retard. La marque autrichienne a en effet quelques projets très attendus qui ont plusieurs mois de retard, et les fans peuvent être très impatients. Aujourd'hui, la roadmap a été mise à jour sur le site de Noctua, et de très belles choses devraient arriver d'ici mi-2026 si tout se passe bien. La fin de cette année sera sous le signe de la gamme chromax.black, et le début de l'année prochaine permettra à Noctua de s'ouvrir à d'autres domaines. […]

Lire la suite

Le jeu fut récemment évoqué lors des festivités organisées à l'occasion des dix ans de la franchise, mais la date de sortie définitive n'était pas encore connue ! Dying Light: The Beast vous promet une nouvelle plongée dans l'effroi à partir du 23 août 2025, les précommandes sont déjà ouvertes, le jeu standard est proposé au prix de 59.99 euros et la version Deluxe est proposée à 69.99 euros, il est rassurant de voir des jeux proposés à moins de 80 euros... Le nouveau titre du studio Techland nous promet un monde ouvert horrifique et oppressant, avec un leitmotiv, la vengeance envers celui qui vous a charcuté, cet apprenti sorcier qui vous a mué en créature hybride mi-zombie, mi-humaine ! De quoi vous offrir certains avantages, mais également de flirter avec la perte de contrôle de vos émotions... […]

Lire la suite

Il y a seulement deux petits jours, nous vous expliquions que les processeurs Ryzen 5 9500F et Ryzen 7 9700F d'AMD étaient visiblement tout proches d'être disponibles, puisque certains revendeurs commençaient à les ajouter à leurs catalogues. Un phénomène qui continue à se propager durant ces derniè...

Switzerland has launched Apertus, a fully open-source, multilingual LLM trained on 15 trillion tokens and over 1,000 languages. "What distinguishes Apertus from many other generative AI systems is its commitment to complete openness," reports CyberInsider. From the report: Unlike popular proprietary models, where users can only interact via APIs or hosted interfaces, Apertus provides open access to its model weights, training datasets, documentation, and even intermediate checkpoints. The source code and all training materials are released under a permissive open-source license that allows commercial use. Since the full training process is documented and reproducible, researchers and watchdogs can audit the data sources, verify compliance with data protection laws, and inspect how the model was trained. Apertus' development explicitly adhered to Swiss data protection and copyright laws, and incorporated retroactive opt-out mechanisms to respect data source preferences. From a privacy perspective, Apertus represents a compelling shift in the AI landscape. The model only uses publicly available data, filtered to exclude personal information and to honor opt-out signals from content sources. This not only aligns with emerging regulatory frameworks like the EU AI Act, but also provides a tangible example of how AI can be both powerful and privacy-respecting. According to ETH Zurich's Imanol Schlag, technical lead of the project at ETH Zurich, Apertus is "built for the public good" and is a demonstration of how AI can be deployed as a public digital infrastructure, much like utilities or transportation. The model is available via Swisscom's Sovereign Swiss AI Platform. It's also available through Hugging Face and the Public AI Inference Utility.

Read more of this story at Slashdot.

Google's September Pixel drop brings the new Material 3 Expressive UI, AI-powered Gboard writing tools, and Bluetooth Auracast upgrades to older Pixel devices, including the Pixel 6 and Pixel Tablet. "Among other tweaks, Google made it possible to add 'Live Effects,' including a few that cover the weather, to your phone's lock screen wallpaper," notes Engadget. "Material 3 Expressive also gives you more control over how the contact cards your phone displays when your friends and family call you look. Even if you're not one to endlessly tweak Android's appearance, as part of the redesign Google has once again reworked the Quick Settings pane in hopes of making it easier to use." On the audio front, Pixel Buds Pro 2 gain intuitive nod-and-shake gesture controls, Adaptive Audio for balanced awareness, and Loud Noise Protection to guard against sudden sound spikes. Voice clarity has also been improved with Gemini Live in noisy environments. A full breakdown of what's new can be found here.

Read more of this story at Slashdot.

The GNOME 49 release candidate "49.rc" was just released as we close in on the stable GNOME 49.0 release in two weeks...

Just days before Apple's expected launch of the satellite-enabled Apple Watch Ultra 3, Garmin unveiled its Fenix 8 Pro -- the company's first smartwatch with built-in inReach satellite and cellular connectivity, SOS features, and a blindingly bright 4,500-nit microLED display. MacRumors reports: With inReach, the Fenix 8 Pro can send location check-ins and text messages over satellite using the Garmin Messenger app. There is also included cellular connectivity, so the smartwatch can make phone calls, send 30-second voice messages, and provide LiveTrack links and weather forecasts when an LTE connection is available. LiveTrack is a feature that allows the wearer's family and friends to keep track of their location during an activity or adventure. For emergencies, there is an SOS feature that will send a message to the Garmin Response center over a satellite or cellular connection. Garmin Response will then communicate with the user, their emergency contacts, and search and rescue organizations to provide help. Garmin says that its Response team has supported over 17,000 inReach incident responses across over 150 countries. The Fenix 8 Pro smartwatch launches September 8, with the AMOLED model starting at $1,200 and the 51mm microLED version priced at $2,000. Both require a paid inReach satellite plan beginning at $7.99 per month for full functionality.

Read more of this story at Slashdot.

An anonymous reader quotes a report from 404 Media, written by Jason Koebler: As I do most nights, I was listening to YouTube videos to fall asleep the other night. Sometime around 3 a.m., I woke up because the video YouTube was autoplaying started going "FEEEEEEEE." The video was called "Boring History for Sleep | How Medieval PEASANTS Survived the Coldest Nights and more." It is two hours long, has 2.3 million views, and, an hour and 15 minutes into the video, the AI-generated voice glitched. "In the end, Anne Boleyn won a kind of immortality. Not through her survival, but through her indelible impact on history. FEEEEEEEEEEEEEEEE," the narrator says in a fake British accent. "By the early 1770s, the American colonies simmered like a pot left too long over a roaring fire," it continued. The video was from a channel I hadn't seen before, called "Sleepless Historian." I took my headphones out, didn't think much of it at the time, rolled over, and fell back asleep. The next night, when I went to pick a new video to fall asleep to, my YouTube homepage was full of videos from Sleepless Historian and several similar-sounding channels like Boring History Bites, History Before Sleep, The Snoozetorian, Historian Sleepy, and Dreamoria. Lots of these videos nominally check the boxes for what I want from something to fall asleep to. Almost all of them are more than three hours long, and they are about things I don't know much about. Some video titles include "Unusual Medieval Cures for Common Illnesses," "The Entire History of the American Frontier," "What It Was Like to Visit a BR0THEL in Pompeii," and "What GETTING WASTED Was Like in Medieval Times." One of the channels has even been livestreaming this "history" 24/7 for weeks. In the daytime, when I was not groggy and half asleep, it quickly became obvious to me that all of these videos are AI generated, and that they are part of a sophisticated and growing AI slop content ecosystem that is flooding YouTube, is drowning out human-made content created by real anthropologists and historians who spend weeks or months researching, fact-checking, scripting, recording, and editing their videos, and are quite literally rewriting history with surface-level, automated drek that the YouTube algorithm delivers to people. YouTube has said it will demonetize or otherwise crack down on "mass produced" videos, but it is not clear whether that has had any sort of impact on the proliferation of AI-generated videos on the platform, and none of the people I spoke to for this article have noticed any change. "It's completely shocking to me," Pete Kelly, who runs the popular History Time YouTube channel, told Koebler in a phone interview. "It used to be enough to spend your entire life researching, writing, narrating, editing, doing all these things to make a video, but now someone can come along and they can do the same thing in a day instead of it taking six months, and the videos are not accurate. The visuals they use are completely inaccurate often. And I'm fearful because this is everywhere." "I absolutely hate it, primarily the fact that they're historically inaccurate," Kelly added. "So it worries me because it's just the same things being regurgitated over and over again. [...] It's worrying to me just for humanity. Not to get too high brow, but it's not good for the state of knowledge in the world. It makes me worry for the future."

Read more of this story at Slashdot.

Tesco is suing Broadcom and reseller Computacenter for at least $134 million, claiming that VMware's perpetual license support agreements were breached after Broadcom's acquisition. The supermarket giant warned it "may not be able to put food on the shelves if the situation goes pear-shaped," writes The Register's Simon Sharwood. From the report: Court documents seen by The Register assert that in January 2021 Tesco acquired perpetual licenses for VMware's vSphere Foundation and Cloud Foundation products, plus subscriptions to Virtzilla's Tanzu products, and agreed a contract for support services and software upgrades that run until 2026. Tesco claims VMware also agreed to give it an option to extend support services for an additional four years. All of this happened before Broadcom acquired VMware and stopped selling support services for software sold under perpetual licenses. Broadcom does sell support to those who sign for its new software subscriptions. The supermarket giant says Broadcom's subscriptions mean it must pay "excessive and inflated prices for virtualization software for which Tesco has already paid," and "is unable any longer to purchase stand-alone Virtualization Support Services for its Perpetually Licensed Software without also having to purchase duplicative subscription-based licenses for those same Software products which it already owns." The complaint also alleges that Tesco's contracts with VMware include eligibility for software upgrades, but that Broadcom won't let the retailer update its perpetual licenses to cover the new Cloud Foundation 9. The filing names Computacenter as a co-defendant as it was the reseller that Tesco relied on for software licenses, and the retailer feels it's breached contracts to supply software at a fixed price. Tesco's filing also mentions Broadcom's patch publication policy, which means users who don't acquire subscriptions can't receive all security updates and don't receive other fixes. The retailer thinks its contracts mean it is entitled to those updates. The filing suggests that lack of support is not just a legal matter, but may have wider implications because VMware software, and support for it "are essential for the operations and resilience of Tesco's business and its ability to supply groceries to consumers across the UK and Republic of Ireland." "VMware Virtualization Software underpins the servers and data systems that enable Tesco's stores and operations to function, hosting approximately 40,000 server workloads and connecting to, by way of illustration, tills in Tesco stores," the filing states. Tesco's filing warns that Broadcom, VMware, and Computacenter are each liable for at least $134 million damages, plus interest, and that the longer the dispute persists the higher damages will climb.

Read more of this story at Slashdot.

Well, here is some sad news... After the better part of two decades at Red Hat, Hans de Goede shared today he will be leaving the company next month. Hans de Goede during his time at Red Hat has been responsible for countless hardware improvements especially for Linux laptops, serves as the x86 platform subsystem lead maintainer for the Linux kernel, and has done immense work over the past 17 years for bettering Linux hardware support especially on consumer devices...

After years of requests, Instagram is finally releasing a dedicated iPad app on September 3rd... "But it will be slightly different than the mobile app users are accustomed to," reports The Verge. From the report: Most significantly, the iPad app will open directly to a feed of Reels, the company's TikTok competitor -- perhaps a sign of the short-form-video times. [...] Other features will be available on iPad: Stories will still line the top of the homepage, and users will be able to switch to a "Following" tab where they'll be able to swipe between feeds that more resemble the mobile Instagram experience (including a chronological option). The bigger screen means more space and fewer clicks: comments on Reels will appear next to full-size videos, and the DMs page will have your inbox alongside chats, similar to what Messenger looks like on desktop.

Read more of this story at Slashdot.

An anonymous reader quotes a report from ZDNet: Over the Labor Day weekend, Cloudflare says it successfully stopped a record-breaking distributed denial-of-service (DDoS) attack that peaked at 11.5 terabits per second (Tbps). This came only a few months after Cloudflare blocked a then all-time high DDoS attack of 7.3 Tbps. This latest attack was almost 60% larger. According to Cloudflare, the assault was the result of a hyper-volumetric User Datagram Protocol (UDP) flood attack that lasted about 35 seconds. During that just more than half-minute attack, it delivered over 5.1 billion packets per second. This attack, Cloudflare reported, came from a combination of several IoT and cloud providers. Although compromised accounts on Google Cloud were a major source, the bulk of the attack originated from other sources. The specific target of this attack has not been publicly disclosed, but we can be sure the intent was to overwhelm the victim's network and render online services inoperative. Cloudflare says its globally distributed, fully autonomous DDoS mitigation network detected and neutralized the threat in real time, without notable impact on customer services or requiring manual intervention. This operation highlights both the rising sophistication of attack methods and the resilience of modern internet infrastructure defenses, especially Cloudflare's use of real-time packet analysis, fingerprinting, and rapid threat intelligence sharing across its network.

Read more of this story at Slashdot.

SUSE's Lubos Kocman announced today on the behalf of the openSUSE team that they will be providing extended maintenance updates and community support with the upcoming Leap 16 release...

U.S. workers are becoming more stressed about their rising personal debts and financial health, a Bank of America survey showed. From a report: Of the respondents polled by BofA, 47% of employed people said they had a sense of financial well-being, dropping from 52% at the start of the year. Nearly 85% of consumers carried some type of personal debt, while 26% of the workforce was seeking help in areas such as emergency savings, paying down debt, and overall financial wellness, compared with 13% in 2023, according to the May survey of more than 1,000 people working full-time.

Read more of this story at Slashdot.

Dolby Vision 2 addresses two widespread TV viewing problems in ways that will likely divide viewers and creators. The format's Content Intelligence feature uses AI and ambient light sensors to brighten notoriously dark content like Game of Thrones' Battle of Winterfell and Apple TV+'s Silo based on room brightness. Authentic Motion grants filmmakers scene-by-scene control over motion smoothing, a feature most cinephiles despise for creating artifacts and making films look like 60fps home videos. Many filmmakers have criticized motion smoothing for undermining artistic intent. Dolby positions the feature as eliminating unwanted judder while maintaining cinematic feel. The format launches in standard and Max tiers for high-end displays.

Read more of this story at Slashdot.

The Rust Foundation today announced the creation of the Rust Innovation Lab that will serve as a stable, neutral home to select Rust projects with governance support, legal and administrative support, fiscal sponsorship, and more...

Ahead of the upcoming Linux 6.18 kernel cycle, which will likely end up being this year's Long Term Support (LTS) version, the AMD Secure AVIC driver appears ready for merging. The AMD Secure AVIC patches were queued this week into a TIP branch and this likely to be submitted for the upcoming Linux 6.18 merge window...

Doucement, mais surement, chaque acteur de l'industrie du semiconducteur se prépare, à son rythme, à l'adoption des EUV High-NA (High Numerical Aperture Extreme Ultraviolet Lithography). Ayant misé beaucoup sur cette évolution technologique, notamment pour son futur procédé 14A, Intel avait été, en...