Jean-Philippe Encausse

Arduino has just introduced three new Modulino modules: the Modulino Hub to support up to eight I2C devices (64 devices with 8 Hubs), the Modulino Extender to extend the I2C range to up to 30 meters, and the Modulino Motors to control two DC motors or a single stepper motor with precision.

I’ll focus on the I2C modules in this post, as I find the Modulino Extender module particularly interesting, since I always considered I2C to be a short-range communication protocol, and never thought of using it for (relatively) long-range wired communication.

Modulino Hub

While the I2C protocol supports up to 127 devices in theory, address conflicts can become an issue in complex builds where multiple sensors or actuators share the same default address. The Modulino Hub aims to solve that, and multiple Hubs can be daisy-chained, each with its own address, to support up to 64 devices.

Modulino Hub specifications:

• Main chip – Texas Instruments TCA9548ARGER 8-channel I2C/SMBus switch
• I2C
  • Address – 0x71 (default); configurable between 0x70-0x77 by jumpers
  • 10x QWICC connectors
    • 8x output connectors for I2C devices
    • 2x input connectors for the Arduino board and another Hub
  • Up to 8x Modulino Hubs can be daisy-chained for a total of 64 devices
  • 4x 4-pin I2C headers (unpopulated) multiplexed with QWICC connectors
  • 4.7kΩ pull-up resistor on all channels
• Power – 3.3V via QWIIC connector or 2-pin header
• Dimensions – 41 x 25.36 mm
• Weight – 4.4 grams

The Modulino Hub is designed for the Arduino UNO R4 WiFi and Arduino UNO Q boards, and you’ll find more technical details and instructions to get started on the documentation website.

Arduino’s I2C Hub module sells for 8.84 Euros/$8.85 on the Arduino store.

Modulino Extender

What’s surprising on the Extender module is that it’s mostly empty with just a single chip (LTC4311) doing all the work, a few passive components, and I2C QWIIC connectors and headers.

 

Modulino Extender specifications:

• Main chip – Linear Technology LTC4311 (LTC4311ISC6) low-voltage I2C/SMBus accelerator
  • Improves the I2C bus rise time transition
  • Ensures data integrity with multiple devices on the I2C bus
  • Improves low-state noise margin
  • Up to 400kHz operation
  • Auto-detect low-power standby mode
  • Low (<5μA) supply current shutdown
  • Does not load the bus when disabled or powered down
  • Strong slew limited pull-up current
  • Wide supply voltage range – 1.6V to 5.5V
  • ±8kV human body model ESD ruggedness
  • Package – 2 x 2mm DFN and SC70 packages
• I2C
  • Address – Not applicable
  • I2C input – QWICC connector and 4-pin header
  • I2C output – QWICC connector and 4-pin header
• Power Supply
  • Voltage – 3.3V via QWICC connector or header
  • Current – ~200 µA typical
• Dimensions – 41 x 25.36 mm
• Weight – 4.4 grams

It’s purely a hardware module, and no specific software or library is needed to use the module. It ships with two QWIIC to female jumper cables (10 cm). The 30-meter range can only be achieved at 100 kHz. The documentation has an important note about the type of cable needed, depending on the range:

• Short runs (< 1 meter) – Standard QWIIC cables work perfectly
• Medium Runs (1-10 meters) – Quality twisted-pair cable, shielded cable recommended
• Long Runs (10-30 meters)
  • Cat5e or Cat6 cable, Shielded Twisted Pair (STP) or Foiled Twisted Pair (FTP)
  • Keep cable runs away from sources of electrical noise
  • Avoid running parallel to power lines

Arduino sells the Modulino Extender for 12.33 Euros/$11.93 on the company’s store.

The post Arduino launches Modulino Hub and Extender modules to support up to 64 I2C devices, 30-meter range appeared first on CNX Software - Embedded Systems News.

Mistral OCR 4 est un modèle de reconnaissance optique de caractères orienté production, capable d’extraire du texte structuré avec bounding boxes, classification de blocs (titres, tableaux, équations) et scores de confiance inline. Il supporte 170 langues, s’intègre comme composant d’ingestion dans des pipelines RAG et enterprise search via le Mistral Search Toolkit, et peut être déployé en self-hosted dans un seul conteneur. Tarification à $4 pour 1 000 pages via l’API, avec 50% de réduction en batch.

---

Commentaires
L'article Mistral OCR 4 : reconnaissance de documents avec bounding boxes et support 170 langues a été posté dans la catégorie Machine Learning de Human Coders News

Formula E is in its final year for the current technical regulations, with a new single-seater EV set to be introduced at the start of next season, which begins in December in Saudi Arabia. The new car, known as GEN4, is a big upgrade—at times more powerful than a Formula 1 car, although heavier and with much less downforce. As speeds rise with the GEN4 car, we knew the sport would become too fast for some of its current venues.

With the release of the season 12 calendar for 2026–2027, that limitation has become clear: a 21-race lineup across 13 cities that now includes several traditional race tracks.

The Saudi double-header is scheduled for December 18 and 19 and is the only season 12 round this year. Then the series starts 2027 off with a string of Formula 1 venues in North America: Mexico on January 16, the Circuit of the Americas in Texas on February 7, and the Miami International Autodrome on February 20. The addition of COTA to Formula E's calendar makes it the seventh US location for the sport since 2015, including the American Airlines Arena in actual Miami; Long Beach, California; Brooklyn, New York; Portland, Oregon; Homestead-Miami, and the Hard Rock Stadium on the outskirts of Miami.

Read full article

Comments

Meta launched a range of cheaper smart glasses featuring the same technology as Ray-Ban Meta but without the branding, and with a compact folding case.

That's not to say that Meta is making these glasses without EssilorLuxottica, the eyewear giant that owns the Ray-Ban brand and that Meta owns a 3% stake of. These new Meta Glasses are still made in partnership with EssilorLuxottica, the two companies confirmed.

Two styles of the cheaper "Meta Glasses", as they're branded, are available: Adventurer and Fury.

Both feature the same chipset, camera, speakers, battery, storage, and connectivity as the Ray-Ban Meta Gen 2 glasses, but are sold starting at $300, an $80 discount compared to the Ray-Bans, owing to their lack of the premium branding.

They also have the more comfortable overextension hinges from the recent Ray-Ban Meta Optics Styles glasses.

Adventurer is a classic rectangular style, offered in 8 colors:

• Classic Black frame with Clear lenses ($300)
• Classic Tortoise frame with Clear lenses ($300)
• Classic Black frame with Brown lenses ($300)
• Classic Tortoise frame with Brown lenses ($300)
• Classic Black frame with Grey Polarized lenses ($330)
• Classic Black frame with Clear to Grey Transitions lenses ($380)
• Merlot frame with Clear to Ruby Transitions lenses ($380)
• Linen frame with Clear to Sapphire Transitions lenses ($380)

Fury is taller and thus more square, offered in 7 colors:

• Classic Black frame with Brown Gradient lenses ($300)
• Mahogany frame with Green lenses ($300)
• Racing Green frame with Light Blue Atlantic lenses ($300)
• Classic Black frame with Grey Polarized lenses ($330)
• Sandstone frame with Dark Amber lenses ($300)
• Classic Black frame with Clear to Grey Transitions lenses ($380)
• Mahogany frame with Clear to Grey Transitions lenses ($380)

We should also point out that the new Meta Glasses are not the cheapest smart glasses the company sells. The Gen 1 Ray-Ban Meta glasses from 2023 are still sold starting at $300 normally, and are currently on sale starting at $224. But the new Meta Glasses have the improvements of the Gen 2: the doubled battery life and 3K video recording.

The Regular Ray-Ban Meta Glasses Get A Gen 2 Upgrade

Gen 2 of the displayless Ray-Ban Meta glasses brings them up to parity with the Oakley Meta HSTN glasses, meaning 3K video recording and longer battery life.

UploadVRDavid Heaney

Flat-Folding Charging Case

The new Meta Glasses range also has a unique advantage over the displayless Ray-Ban and Oakley branded smart glasses. They feature a folding charging case that seems to be very similar to the case of Meta Ray-Ban Display.

0:00

/0:31

1×

As we noted in our Meta Ray-Ban Display review, the advantage of a folding case is that, while you're wearing the glasses, the case can fit into a jacket pocket, or even a particularly deep trouser pocket.

Kylie Jenner Glasses & Charging Stand

The cheaper glasses are not Meta's only new smart product launch this week.

The company also worked with Kylie Jenner on a completely new smart glasses style called Starfire, which has a notable resemblance to the upcoming Gentle Monster smart glasses powered by Google.

Meta Preempts Google’s Gentle Monster Glasses With Kylie Jenner ‘Starfire’ Design

Meta teamed up with Kylie Jenner for new ‘Starfire’ smart glasses featuring her voice for Meta AI, a gemstone in the lens, and a metal nose pad to prevent absorbing makeup.

UploadVRDavid Heaney

Further, for people who want to keep their smart glasses available by their bedside, Meta launched a $60 Charging Stand, made of stainless steel, with a USB-C port. It supports the company's displayless Ray-Ban and Oakley HSTN smart glasses, charging them to 50% in 20 minutes and fully within an hour. Meta Ray-Ban Display and Oakley Meta Vanguard are not supported.

Meta Now Sells A Charging Stand For (Most Of) Its Smart Glasses

Meta now sells a $60 stainless steel Charging Stand compatible with most of its smart glasses, providing 50% charge in 20 minutes or full charge in 1 hour.

UploadVRDavid Heaney

OpenC6 BIOS is an open-source project by Rompass that takes a different approach to MCU development. It adds a BIOS-like system to the ESP32-C6, so the system part and application code can run separately rather than being combined into a single firmware image.

In traditional development, hardware setup, networking, and application code are combined into a single firmware image. But the OpenC6 BIOS operates differently: it runs as the base system on the ESP32-C6 and handles hardware initialization and system functions. Instead of flashing a full application each time, it can load small payload programs separately. These payloads can run from RAM or flash (XIP) and use system functions through a simple Application Binary Interface (ABI), without needing the full firmware. In some ways, it operates like AkiraOS, but with a BIOS-like interface.

Key features of the OpenC6 BIOS architecture:

• Dynamic CPU scaling – Adjusts CPU frequency based on load using FreeRTOS idle time (80, 120, and 160 MHz)
• Modular firmware architecture – Hardware setup is handled by the BIOS, while applications are compiled as small independent payloads (2–10 KB)
• System Call Interface (ABI) – Payloads can access Wi-Fi, math, and crypto functions provided by the BIOS without linking to the full ESP-IDF
• LP-Core management engine – Uses the ESP32-C6 low-power RISC-V core to monitor system health, watchdog timers, and power events independently of the main application
• Network boot (PXE-like) – Payloads can be downloaded over Wi-Fi and executed without flashing via USB
• A/B OTA updates – Firmware updates include rollback support to avoid bricking the device
• Web-based setup interface – A local Wi-Fi access point (BIOS_SETUP_C6) hosts a browser-based configuration page (192.168.4.1) similar to a PC BIOS setup utility with a 1280×1024 blue-screen look
• Aura Sync RGB – Supports Aura Sync RGB Flow and POST diagnostic codes
• Misc
  • Power Button – GPIO 4  (Short click: Boot; Hold 3s: Setup; Hold 5s: Hard Reset)
  • BOOT Button – GPIO 9 (Standard BOOT pin. Hold during startup for Boot Menu)
  • POST LED – GPIO 8 (Addressable RGB for diagnostics and Aura Sync)
  • Clear CMOS – GPIO 2 (Short during boot to reset NVRAM to factory defaults)
  • Payload RX / TX – GPIO 18 / GPIO 19 (Dedicated UART for the Serial Bootloader)

The BIOS runs first after boot and performs hardware initialization. It then exposes services through a defined ABI, so compiled payloads can be uploaded to a UART serial loader, network-booted over an HTTP server, or executed directly from flash (XIP), without using the full ESP-IDF.

This is similar to old PC BIOS, which gave basic functions before starting the OS, but here it is used to run small embedded programs instead of a full operating system.

The project is designed for generic ESP32-C6 boards with at least 8MB of Flash. The ESP32-C6 Zero board was used for development and testing, but the OpenC6 BIOS project should also work with boards such as the Olimex ESP32-C6-DevKit-Lipo, WeAct ESP32-C6 (8MB), Arduino Nesso N1, and others. Modifications for each board may be required, since the firmware assigns GPIOs for power and reset control, an RGB LED for status indication, a UART for loading payloads, and configuration pins similar to CMOS reset pins.

To get started, you need to install ESP-IDF (development version) and build the BIOS firmware. Payloads are compiled separately using the provided tools available in the tools directory.  Network booting is supported by running a simple local HTTP server and pointing the device to the payload URL through the web setup interface.

OpenC6 BIOS is open source under the MIT license. The code, documentation, and payload build instructions are available on GitHub. The developer is also looking for contributors to work on a custom zRAM and a lightweight file system for RISC-V.

Via Hackaday

The post OpenC6 BIOS project brings PC-like firmware to ESP32-C6 MCU with network boot and OTA update support appeared first on CNX Software - Embedded Systems News.

Sonocotta’s Louder-ESP32-Mini is a compact version of the Louder ESP32 board that converts old speakers into WiFi (and Bluetooth) connected speakers for your Smart Home.

The board aims to replace the speaker terminals on your old speakers, and comes in two sizes: 42x42mm for small speakers, and 52x52mm for larger models. It’s powered by an ESP32-S3 wireless SoC with 8MB PSRAM and equipped with a high-quality TAS5805M DAC with DSP features.

Louder-ESP32-Mini specifications:

• Wireless Module – ESP32-S3-WROOM-N8R8
  • SoC – Espressif Systems ESP32-S3
    • CPU – Dual-core Tensilica LX7 up to 240 MHz with vector extension for AI/ML workloads
    • RAM – 512KB SRAM, 8MB PSRAM
    • Storage – 8MB flash
    • Wireless – WiFi 4 and Bluetooth LE 5
  • Antenna – PCB antenna
• Audio
  • DAC – Stereo TAS5805M I2S DAC with built-in DSP and D-Class amp
  • DSP Features – EQ, FIR, DRC, AGL, etc…
  • Speaker output via two WAGO connectors
    • 4Ω, 1% THD+N
      • 42×42 mm model – 1x 5W
      • 52×52 mm model – 1x 15W
    • 8Ω, 1% THD+N
      • 42×42 mm model – 1x 3W
      • 52×52 mm model – 1x 8W
• Misc – Reset and user buttons
• Power Supply
  • 42×42 mm model – 5V via USB-C port
  • 52×52 mm model – 12V via USB-C port (USB PD)
• Dimensions – 42×42 mm or 52×52 mm

As I understand it, you’d remove the standard speaker terminal cup shown in the foreground above, connect the internal wires to the WAGO connector of the Louder ESP32 Mini board, and insert it in the opening before securing it with the four screws from the original cup. Sonacotta warns users that Louder boards will connect to passive speakers only, and headphones or an external amplifier can be used (they have other products for that).

The board supports several firmware options, including Squeezelite-ESP32 for Spotify Connect, AirPlay, Bluetooth, or Logitech Media Server, Snapclient multi-room audio streaming with snapserver, and ESPHome for Home Assistant integration, although you could also roll out your own firmware. It’s basically the same as for other audio boards from the company, and you’ll find details on GitHub and the documentation website. The project is open-source hardware, and the EasyEDA schematics, Gerber files, and other files can be found on the same GitHub repo.

The Louder ESP32 Mini board can be purchased for $15 (42x42mm) or $20 (52x52mm) on Lectronz or Tindie.

The post Louder ESP32 Mini board adds WiFi and Bluetooth to old speakers for SqueezeLite, Snapclient, or ESPHome support appeared first on CNX Software - Embedded Systems News.

Decrazyo vient de sortir ANES , un mod hardware qui transforme votre vieille NES en lui collant un deuxième PPU, la puce graphique de la console.

Résultat, plus de couleurs à l'écran, plus de sprites, et même du parallax scrolling, ces décors qui défilent sur plusieurs plans comme sur les bornes d'arcade. Que des trucs que la petite Nintendo de 1985 n'a jamais su faire, la miskine.

Makerfabs MaTouch ESP32-P4 TFTTouch 10.1 MIPI with 4G LTE is an ESP32-P4 + ESP32-C6-based display development platform for HMI edge AI, and connected embedded applications. The development board features a 10.1-inch IPS LCD, very similar to those of Seeed Studio’s reTerminal D1001, Waveshare ESP32-P4-WIFI6-Touch-LCD, and CrowPanel Advanced 7-inch displays.

However, what makes it unique compared to other devices is that it also features a SIM7670G 4G LTE modem with a SIM card slot, an RJ45 Ethernet jack, and a 2MP camera. The HMI display also comes with microphone and speaker connectors, a microSD card slot, various USB interfaces, lithium battery support, and two 2.54 mm pitch GPIO expansion headers exposing UART, I2C, SPI, and other peripheral interfaces. All these features make it suitable for HMI, edge AI, and connected embedded applications.

Makerfabs MaTouch ESP32-P4 10.1-inch 4G LTE specifications:

• SoC – Espressif Systems ESP32-P4NRW32
  • CPU
    • Dual-core 32-bit RISC-V HP (High-performance) CPU @ up to 400 MHz with AI instructions extension and single-precision FPU
    • Single-RISC-V LP (Low-power) MCU core @ up to 40 MHz
  • Memory
    • 768 KB HP L2MEM (for dual-core CPU), 32 KB LP SRAM, 8 KB TCM (for LP MCU core)
    • 32MB PSRAM
  • Storage – 128 KB HP ROM, 16 KB LP ROM
  • GPU – 2D Pixel Processing Accelerator (PPA)
  • VPU – H.264 and JPEG codecs support
• Storage
  • 16MB QSPI NOR Flash
  • MicroSD card slot
• Display
  • 10.1-inch IPS LCD, 1280×800 resolution
  • MIPI DSI interface (JD9365 LCD driver)
  • 10-point capacitive touchscreen (GSL3680 touch driver)
• Camera – 2MP MIPI camera (SC2336) supporting up to 1920×1080 resolution
• Audio – Onboard microphone, speaker output header via ES8311 audio codec and NS4150 audio amplifier
• Networking
  • Wireless –  ESP32-C6
    • 2.4 GHz WiFi 6
    • Bluetooth 5.x LE
    • 802.15.4 radio for Zigbee, Thread, and Matter
  • Wired Networking – 10/100Mbps Ethernet via IP101GRI PHY
  • Cellular
    • SIMCom SIM7670G 4G LTE Cat 1 module with Nano SIM card slot
    • 4G LTE antenna
• USB – 6x USB Type-C ports
  • USB-TTL (CH340K via USB to UART chipset)
  • USB-NATIVE
  • USB-CDC/JTAG
  • USB-OTG
  • USB-DEVICE
  • USB-LTE (dedicated to the SIM7670G cellular module)
• Expansion – 2x 40-pin headers exposing 26x GPIOs, 3.3V, 5V, and GND
• Misc
  • Power ON/OFF switch
  • Reset and Boot buttons
  • Power LED, Status RGB LED
  • PCF8563T RTC with CR1220 battery support for timed interrupt wake-up
• Power Supply
  • 5.0V (4.0V~5.25V) via USB Type-C port
  • 2,500 mAh battery (2-pin lithium battery connector with TP4056 charging IC)
• Dimensions – TBD

On the software side, Makerfabs mentions support for the ESP-IDF framework, the Arduino IDE, and PlatformIO, as well as the LVGL graphics library for designing complex, fluid user interfaces. The new 4G LTE ESP32-P4 display shares the same GitHub repository as the earlier “Matouch_ESP32-P4_TFT_with_Touch_10_1” without cellular connectivity. We could only find generic ESP-IDF samples and PDF schematics there, so if you want to use Arduino, you’re on your own… The documentation is also shared between the two displays; it includes a 4G LTE HTTP section explaining how to send AT commands to control the modem with the ESPBrookesia firmware.

The MaTouch ESP32-P4 10.1-inch MIPI HMI display development board with 4G LTE is sold on the Makerfabs store for $79.80. The board comes pre-installed in a basic enclosure, and the kit includes the 32GB MicroSD card.

The post Makerfabs MaTouch ESP32-P4 10.1-inch HMI display features 4G LTE and Ethernet connectivity, a 2MP camera appeared first on CNX Software - Embedded Systems News.

Misjudge a curb or miss a step on the stairs, and there is a split second of panic as your foot doesn't land when you expect it to. That brief loss of pressure can be enough to throw off your balance entirely.

While mouse sensors have improved a lot over time, the scroll wheel has mostly stayed the same. Normal scrolling requires you to keep flicking your finger up and down, which can get uncomfortable during long sessions like coding or video editing. To change this, Airra Labs has introduced the Rotary Mouse, which uses a rotary wheel in place of the scroll wheel to let you scroll with a simple circular motion.

It comes in two variants: a standard and a vertical version. The standard variant features a larger circular, jog-wheel-like dial on top that lets you scroll by rotating your finger. The vertical variant also comes with the same wheel but is designed so you can scroll like a regular mouse by flicking your finger up and down, while still using the same circular motion. This provides a wider range of movement and can make scrolling faster and easier than with a regular mouse.

Rotary mouse specifications:

• Variants – Standard and Vertical
• Connectivity – 2.4 GHz wireless connection via included USB wireless receiver
• Sensor – Optical sensor with adjustable sensitivity (800 / 1200 / 1600 DPI)
• Buttons
  • Silent left and right buttons, plus an integrated middle click button on the rotary wheel
  • Rotary wheel supporting continuous rotation with tactile feedback
• Power – AAA alkaline battery (included)
• Dimensions – 119  × 64 × 40 mm
• Weight – 59 grams excluding the battery
• Color – Matte black

The mouse uses continuous rotation with tactile feedback (detents) for control, which enables up to 2.5× faster scrolling according to the company, while maintaining fine, pixel-level precision. The device can also function as a jog wheel or analog-style steering input for various applications and connects via a standard 2.4 GHz USB dongle.

With all these features, it looks like the hardware is primarily designed for developers, video editors, and simulation gamers. It helps developers and data analysts quickly scroll through large code files, JSON data, and spreadsheets. For video editing, it works like a jog wheel, making it easier to scrub through timelines in tools like DaVinci Resolve, Adobe Premiere Pro, and Final Cut Pro. It can also be used in driving games like Euro Truck Simulator 2 and Assetto Corsa, where the dial works as a simple steering control. I might also be useful to prevent repetitive strain injuries (RSI).

In the past, we have written about various types of mice, including the Sanwa Supply RF mouse, the Ploopy Bean open-source hardware pointing stick mouse, and the Pepper Jobs W10 GYRO Air Mouse, all of which are unique in their own way.

The Rotary Mouse is available on Kickstarter, where the  Super Early Bird ($47) tier is already sold out, but you can still get the standard version for around $76 (with preset settings) or $86 for the second-batch version. The vertical ergonomic version is priced at about $105, and all the varieties include free worldwide shipping, with the first batch expected in November 2026 and the second batch in January 2027.

The post Rotary Mouse puts a rotary wheel into a standard desktop mouse (Crowdfunding) appeared first on CNX Software - Embedded Systems News.

Holographic displays sound very fancy but you can build various simple types yourself at home. [Julius Makes] whipped up a neat design that shows a different image depending on the position from which you view it. 

Running the show is a Wemos D1 devboard equipped with the ESP8266 microcontroller. It’s hooked up to a pair of OLED displays over I2C. The displays are placed in a 3D printed assembly that aims each one at a beam-splitter cube. This bounces light projected into one face through 90 degrees, and out another face. By leveraging this, it’s possible to aim each display at one face and bounce it out another, such that looking at either side of the beamsplitter cube shows a different image. Since the beamsplitter cube also allows some light to be transmitted directly through as well, the image from each display appears to float in space.

[Julius] notes that this setup is being used in a puzzle box game, while wondering whether there’s any other fun ways to leverage this technique. We’ve seen some other neat holographic displays before, too, like this neat Holochess build.

UC Berkeley's electric nose uses a carbon nanotube gas sensor chip to detect spoiled food and allergens, but crowded refrigerators and mixed-food smells are still the real test.

Le FBI possède sa propre ville, sauf que personne n'y habite, et pour cause, elle a été montée de toutes pièces dans un hangar de Huntsville, en Alabama, avec ses maisons meublées, son hôtel, sa station-service, son épicerie, son tribunal, son hôpital et jusqu'à sa compagnie d'électricité, le tout dans un seul but assez vertigineux, la pirater dans tous les sens sans jamais déranger âme qui vive.

Le décor porte d'ailleurs un nom, le Kinetic Cyber Range, près de 2 000 mètres carrés de fausse bourgade américaine ouverte en février 2025 et pensée comme un gigantesque bac à sable pour cyberattaques en conditions réelles.

Kyiv peut désormais activer l'assistance d'urgence européenne face aux cyberattaques de grande ampleur. Une décision symbolique autant que stratégique.

Vous voulez savoir comment GPT-5.5, Claude ou Grok raisonnent quand on les met sous pression ? Hé bien filez-leur un jeu de stratégie et regardez-les se faire la guerre.

C'est tout le principe d' Age of LLM , monté par Rymentz, un lecteur du blog où deux IA s'affrontent au tour par tour, sans aucun humain aux commandes.

C'est un benchmark 1v1 dans lequel on balance deux modèles sur la même carte, on leur donne les règles et puis c'est tout, sans aucun autre conseil de stratégie. À eux alors de déduire seuls comment contrer les attaques, quand bâtir leur économie, quand lever une armée pour aller défoncer la base d'en face, quand négocier la paix et quand carrément appuyer sur le gros bouton rouge. Parce que dans ce benchmark, on peut gagner de trois façons : la conquête militaire, la victoire diplomatique, ou la bombe nucléaire.

Tout est ensuite rendu public sur le site, avec un classement ELO et le replay de chaque partie. Pour le moment, 13 modèles y sont classés, de GPT-5.5 à Grok en passant par Claude, Gemini, DeepSeek ou Qwen.

Sur les 43 parties jouées par la version 0.12.0 du moteur, 38 se terminent par un champignon atomique. C'est énorme je trouve... La diplomatie ne l'a emporté que 2 fois, et le militaire 3 fois. Mettez deux de ces modèles face à face avec une bombe à portée de main, et neuf fois sur dix, ça finit donc par appuyer sur le bouton rouge. Elles nous ressemblent tellement ces IA quand même...

Des chercheurs qui ont fait jouer des LLMs à des simulations de crise nucléaire ont relevé une tendance comparable. Quand l'option d'escalade existe, les modèles finissent le plus souvent par la prendre. C'est pour ça qui ne faut pas confier d'armes à des IA les amis... Car elles s'en serviront c'est certain !

Le jour où j'ai compris que les IA et les armes, c'était pas un bon mélange à faire

Et chacun de ces modèles a vraiment son style... Y'a qu'à regarder le classement qui est sans pitié. GPT-5.5 écrase tout le monde, 6 victoires sur 6, 100% de winrate, premier à l'ELO... mais il médite presque 5 minutes par tour avant de jouer.

À l'autre bout du tableau, on retrouve Grok 4.3 qui finit bon dernier, zéro victoire en 4 parties, mais qui joue le plus vite de tous, avec 7 secondes par tour (autant dire qu'il ne réfléchit pas trop).

Les Claude, eux, jouent propre et solide mais sans jamais dominer. Ça me rappelle l'IA qui galérait à finir un niveau de Super Mario ...

Mais le chiffre le plus parlant, c'est le taux de coups illégaux, c'est à dire le pourcentage de fois où un modèle propose une action interdite par les règles.

Par exemple, sous pression, Grok 4.3 culmine à 8,6% d'actions de triche, soit le pire du lot. Cela veut dire que sa capacité à suivre des règles sous pression est moins bonne que celle des autres modèles.

Ça résonne forcement avec ce papier d'Adrian de Wynter dont je vous parlais récemment , celui qui montrait qu'Age of Empires II a autant d'attributs "humains" que votre ChatGPT, histoire de dégonfler la hype sur les IA soi-disant conscientes.

Age of LLM, lui, prend le problème par l'autre bout, où au lieu d'analyser on fait jouer et on regarde ce qui sort. Derrière ce projet, on retrouve le même lecteur qui avait bricolé Nuclear Arms Race il y a quelques années et dont je vous avais parlé. Il a repris le concept, sauf que cette fois ce sont des IA qui jouent entre elles.

En tout cas, comme spectacle, c'est très hypnotique. Lancez un replay, vous allez voir... En regardant deux IA se bluffer et se faire exploser la tronche vous permettra d'en apprendre plus sur leur raisonnement qu'avec dix pages de benchmark.

Des ingénieurs de l'université du Texas à Austin ont mis au point un vêtement qui tire de l'eau potable directement de l'air ambiant, sans réservoir à remplir ni source à proximité, simplement grâce au tissu dont il est fait. Le travail, mené par Guihua Yu et son équipe de la Cockrell School of Engineering, avec le chimiste Keith Johnston du département de génie chimique McKetta et l'auteur principal Weixin Guan, vient d'être publié dans la revue scientifique Science Advances.

Le secret tient dans la matière. Il s'agit d'un tissu en hydrogel, c'est-à-dire un gel ultra-absorbant ici fabriqué à partir de matières issues de la biomasse, donc d'origine végétale, qui capte les molécules d'eau présentes dans l'air sous forme de vapeur avant de les faire passer à l'état liquide sur la surface des fibres puis de les stocker en profondeur.

Toute la prouesse réside dans la vitesse à laquelle l'eau franchit ces étapes, l'équipe ayant conçu un chemin qui la fait circuler très rapidement de la vapeur vers le liquide puis vers le textile, ce qui lui vaut une efficacité de trois à dix fois supérieure aux matériaux de récupération atmosphérique habituels.

Une fois l'humidité captée, elle est dirigée vers de petits modules détachables que l'on retire pour les glisser dans un collecteur pliable, où la chaleur du soleil libère l'eau emprisonnée dans le gel, qui se condense alors et devient buvable.

Côté rendement, la veste produit entre 400 et 900 millilitres d'eau potable par jour, soit grosso modo une à deux petites bouteilles, selon le taux d'humidité de l'air. Une version posée au sol plutôt que portée grimpe, elle, jusqu'à 1,3 litre quotidien, et tient ce chiffre aussi bien dans un climat aride que dans une ambiance tropicale moite.

La comparaison avec Dune était écrite d'avance. Dans le roman de Frank Herbert porté à l'écran par Denis Villeneuve, les habitants de la planète désertique Arrakis enfilent des stillsuits, ces combinaisons qui recyclent la transpiration et l'urine de celui qui les porte pour ne gaspiller aucune goutte d'eau dans un monde où elle vaut plus cher que tout. Ici, c'est nettement plus ragoûtant, puisque la veste ne prélève rien sur le corps et se contente d'aspirer l'humidité de l'air.

Et le projet ne s'arrête pas à un simple blouson, l'équipe imaginant déjà glisser le même tissu dans des sacs à dos, des tentes ou des abris d'urgence, histoire de transformer du matériel qu'on transporte déjà partout en petite usine à eau qui ne réclame ni pile ni branchement électrique.

Bon, il reste des inconnues de taille, puisqu'on ignore le poids réel du vêtement, son prix éventuel et surtout sa tenue dans le temps une fois enchaînés les cycles d'absorption et de chauffe. Et 900 millilitres, c'est encore loin de couvrir les besoins d'un adulte qui marche toute la journée en plein cagnard.

N'empêche, pour un randonneur perdu ou une région privée d'eau, tirer un demi-litre du ciel sans la moindre nappe à proximité, ça peut tout changer.

Source : Techspot

7 dollars, c'est le prix d'une bague connectée chinoise sur AliExpress, et c'est surtout tout ce que Saksham Bhutani a payé pour se bricoler un coach santé privé. Son truc s'appelle PulseLoop , et c'est une app iPhone open source qui transforme ce gadget à deux balles en tracker de fréquence cardiaque, de sommeil et d'activité, sans abonnement ni cloud à la con.

Le principe, c'est de prendre une bague avec capteurs vendue trois fois rien (la même qui traîne sur AliExpress sous le nom de SMART_RING et en un peu plus cher sur Amazon ) et de la débrancher complètement de son app d'origine, la fameuse JRING. À la place, vous clonez le projet, vous compilez ça dans Xcode, et hop, votre iPhone causera alors directement à la bague en Bluetooth.

Fréquence cardiaque, nombre de pas, calories brûlées et kilomètres sont relevés par la bague, ainsi que le parcours par le GPS du téléphone (si vous avez votre smartphone avec le GPS actif dans la poche), et tout est agrégé et stocké en local sur le téléphone et nulle part ailleurs.

Côté sommeil, on peut grâce à ça, suivre jour après jour la qualité de la nuit qu'on vient de passer et ainsi essayer d'atteindre plus de régularité pour améliorer son score de sommeil. Pour ma part, ce genre de trackers d'activité ne sont jamais restés bien longtemps à mon poignet, mais j'avoue que bidouiller le sien change un peu la donne et me fait reconsidérer la chose.

Et là où ça devient du vrai boulot de bidouilleur, c'est que rien de tout ça n'était documenté. Saksham a donc sorti Wireshark et un dongle Nordic nRF52840 pour renifler les échanges Bluetooth de la bague, puis il a tout reconstruit. Résultat, un second repo qui documente le protocole de A à Z, avec une petite CLI Python où vous tapez des commandes du genre hr run 45 ou spo2 run 45 pour déclencher une mesure. Y'a même un mode selfie planqué dedans, vous serrez le poing et la bague déclenche la prise de photo sur le téléphone !

Le truc marrant (enfin, marrant... je me comprends), c'est ce qu'il a trouvé en chemin... Car OUI la bague balance tout en clair sur un service maison avec 0 chiffrement. ou signature. Vos battements de cœur transitent donc à poil sur le Bluetooth ! D'ailleurs, on a vu avec l'affaire Strava que ces données de tracking peuvent révéler bien plus que prévu, alors une bague à 7 balles qui cause sans aucune protection, ça craint un peu...

En tout cas, c'est ça qui a énervé Saksham au départ. Il avait acheté un Google Fitbit, il aimait l'idée, mais pas le reste. A 100 balles l'appareil + 10 balles de plus par mois d'abonnement, et une app bourrée d'IA à tous les étages, sans oublier ce flux continu de son rythme cardiaque, de son sommeil et de ses déplacements envoyé non-stop chez Google pour qu'ensuite ils les revendent... Ça commençait à faire un peu trop pour lui. Chez Whoop, Oura ou d'autres, c'est tout pareil, vos données SONT le produit.

C'est pour cela qu'avec PulseLoop il a décidé de casser ça...

Au-dessus, il a même collé un coach IA plutôt malin qui lit vos mesures en local et vous sort des trucs du style "ton rythme au repos a grimpé les trois nuits où tu as dormi moins de 6 heures". Il fait des graphiques, repère des tendances, retient vos objectifs. Après ce coach-là tourne sur l'API d'OpenAI avec VOTRE clé du coup dès que vous lui posez une question, vos données partent chez OpenAI et vous payez à l'usage. Mais bon, même sans ça, le reste de l'app marche très bien.

Après, faut pas rêver sur la qualité de l'appareil car une bague à 7 dollars c'est pas un capteur Oura... les capteurs valent ce qu'ils valent et le décodeur de sommeil est encore expérimental (il ne choppe même pas les phases REM pour l'instant). On est dans l'esprit d' AsteroidOS , sauf qu'ici on "libère" une bague plutôt qu'une montre.

Mais bon je me dis que les capteurs vont évoluer, ils vont en devenir de mieux en mieux et cet outil sous licence libre sera toujours là pour en reprendre le contrôle. Pour moi c'est un projet à suivre, même s'il est encore ultra jeune (et uniquement sur iPhone pour le moment).

CrankGPT vous connaissez ? Elle fait tourner un assistant vocal complet, reconnaissance de la voix comprise, sans prise murale, sans batterie et sans serveur distant, et pour l'alimenter vous tournez une manivelle dont la résistance grimpe quand le modèle réfléchit.

Derrière, deux anciens de Google. Katrin Tomanek, informaticienne, et Alex Kauffmann, passé par le laboratoire ATAP, la division des projets un peu fous. Ils ont monté Squeez Labs ensemble.

The Snapmaker U1 3D printer is an impressive machine for the price, but [Beaver Works] found the optional factory-offered top cover a wee bit expensive for his tastes. The solution? 3D print a fixture and use a clear 45 L Samla storage box from IKEA as an effective and affordable cover for the machine.

Why a cover?  A cover helps retain heat and block drafts, which can help improve print quality. A cover also keeps the machine’s insides dust and debris-free, not to mention serving as a decent barrier to curious fingers or paws.

This is a great use of an off-the-shelf product that performs at least as well as any bespoke solution. The nature of printer enclosures makes them trickier than one might think, with the size and weight of materials often driving costs up for something that seems relatively simple in concept. Getting one by 3D printing the fixtures and purchasing the bulky part locally and affordably is a great alternative. IKEA even sells the box’s lid separately, so one can buy just the box and isn’t stuck with an unused lid afterward.

Integrating off-the-shelf components into a design is often risky because much of it is outside the designer’s control. Availability can change, and a manufacturer might alter dimensions or design elements without any notice. But IKEA’s storage products are pretty well standardized and work really well for this purpose.

On the off chance you need a design tweak, [Beaver Works] has provided STEP files for the 3D-printed parts, something we always love to see.

Des chercheurs new-yorkais ont pris des embryons humains sains, et ont réussi à modifier leur code génétique en se concentrant sur une seule lettre. L'outil utilisé, appelé édition de base,...

Philips Skylight is a slim ceiling panel designed to recreate the brightness, depth, and rhythm of natural daylight inside your home.

• NatureConnect Daylight Simulation: Using Signify’s NatureConnect technology, the panel reproduces natural light shades and depth. It gives you the visual feel of a sky-facing window rather than a standard ceiling light.
• Auto Day Rhythm: Brightness and color tone shift automatically throughout the day, delivering brighter, blue-enriched light in the morning and warmer tones in the evening.
• Five Preset Scenes: Users can choose from five daylight-inspired lighting scenes to match different activities.
• VitaUp UV-B Module: Select variants include an integrated UV-B module. It supports the body’s natural vitamin D production indoors, with an automatic shutoff after eight hours for added safety.
• IP44 Rating: All variants carry an IP44 rating, making them suitable for bathrooms and other damp environments.

Philips Skylight suits anyone spending long hours indoors who wants light that feels closer to the real thing.

The post Philips Skylight appeared first on Gadget Flow.

Capter la vapeur d'eau dans l'air pour en faire une ressource buvable, l'idée existe depuis plusieurs années. Mais jusqu'ici, les dispositifs capables de le faire restaient encombrants : boîtiers fixes, panneaux sorbants, installations lourdes à déplacer... L'équipe menée par Guihua Yu, professeur titulaire au département de génie mécanique de la Coc...

Détourer un sujet filmé sur un fond vert, c'est le cauchemar classique du compositing. Entre les cheveux rebelles, le flou de mouvement et cette satanée frange verte sur les bords, on y passe vite des heures. Faut dire que la plupart des outils de détourage automatiques actuels c'est de la merde, puisqu'ils se contentent d'extraire des masques binaires un peu brutaux qui massacrent les détails fins.

Et c'est là qu'intervient CorridorKey , un outil dont le code est disponible sur GitHub et développé par Niko Pueringer de l'équipe Corridor Crew . Au lieu de bêtement découper la silhouette, ce réseau de neurones fait un travail de dé-mélange de couleurs (unmixing) hyper précis. Cela lui permet de reconstruire la couleur d'origine de chaque pixel en retirant le fond vert, tout en générant un canal alpha linéaire très propre... et c'est dingue parce que ça change tout en terme de qualité de rendu !

L'interface de EZ-CorridorKey après traitement

Pour les pros du VFX qui bossent sur Nuke, Resolve ou Fusion, l'outil gère nativement le format EXR en 16 et 32-bit linéaire. Ça conserve également toute la dynamique des couleurs pour l'intégration finale. Du coup, ça vaut le coup d'oeil si vous voulez intégrer ça proprement. D'ailleurs au passage, si vous cherchez un éditeur libre plus classique, jetez un oeil à Shotcut , je l'avais testé y'a pas longtemps et c'est vraiment cool.

Maintenant, attention avec CorridorKey car si traiter des images pour vos projets commerciaux est autorisé, la revente, les API payantes ou l'intégration logicielle ne marchent pas sans accord écrit préalable. Pour les repères de détourage (les alpha hints), je préfère BiRefNet car c'est plus léger mais autrement, le système s'appuie sur des modèles comme Generative Video Matting (GVM) ou VideoMaMa qui sont super quand même.

Côté matériel, Niko a développé le projet avec une NVIDIA RTX Pro 6000 de 96 Go de VRAM (en gros, une machine de guerre). Mais heureusement, la communauté qui a moins de sous a optimisé tout ça. Du coup, la version actuelle tourne aujourd'hui sur des GPU de 6 à 8 Go de VRAM, mais n'oubliez pas qu'il faut CUDA 12.8 sous Windows pour faire tourner le bouzin.

Comparaison de qualité de détourage

Les utilisateurs de Mac ne sont pas oubliés non plus puisque le soft tourne aussi sur les puces Apple Silicon M1+ grâce à la mémoire unifiée. L'installation est facilitée également par l'usage du gestionnaire de paquets uv, qui s'occupe d'installer Python, de créer l'environnement virtuel et de récupérer les dépendances sans que vous ayez à lever le petit doigt.

Et si la ligne de commande vous donne des boutons, il existe un fork fort sympathique baptisé EZ-CorridorKey . Ça ajoute une interface graphique simple, ce qui rend l'outil accessible aux artistes qui préfèrent cliquer sur des boutons. Et si vous aimez les effets analogiques, vous pouvez aussi le coupler avec ntsc-rs pour donner un look rétro à vos vidéos.

Bref, CorridorKey est un excellent moyen de tester la puissance de l'IA appliquée au détourage VFX pro sans se ruiner en logiciels propriétaires.

A vous de jouer maintenant !

Google DeepMind has unveiled TacticAI, an AI system developed alongside Liverpool FC that can predict football player movement up to eight seconds in advance while also generating tactical recommendations for set pieces.

Among the ways Formula 1 has changed in the 21st century has been its adoption of driver-in-the-loop simulators. It all started in the early 2000s, probably at McLaren, maybe at Toyota or Ferrari; F1 teams are notoriously secretive about their performance advantages. Along the years, they've gotten more and more capable, but so too have high-end consumer sims like the multi-axis setups that cost tens of thousands of dollars. What is it that makes the multimillion-dollar simulators used in F1 that much more expensive, and that much better for the job?

For one thing, latency.

"There's this intimate link between the inputs that [a driver] provides to the car, the way the car responds, and then the driver immediately feels that and reacts to it. So this is a very dynamic closed loop involving the driver and the car," explained Ash Warne, founder and CTO of Dynisma Motion Generators, a UK-based simulator company that supplies Ferrari, Alpine, and soon Cadillac with DiL simulators that can cost as much as $10 million.

Read full article

Comments

La Belgique a-t-elle vraiment approuvé le FSD de Tesla sur son territoire ? Alors que le constructeur n'a pas officialisé l'homologation, la ministre des Transports de Flandre a fait savoir à tout le monde qu'elle avait donné son accord.

What would you see if you tried to travel alongside a light wave at the speed of light? And suppose you held a mirror in front of you as you zipped along. What would you see in the mirror?

This and similar thought experiments were posed by the young Albert Einstein to himself in his teens. It’s come to be known as Einstein’s Mirror and is also the title of a popular book on relativity.

It would at first seem that light, reflected off your face, could never reach the mirror to, in turn, reflect back into your eyes to see it. So what would you see?

It was only years later that Einstein developed a theory that answered this puzzle. And it required some fundamental adjustments to how we understood the world, which still bend my mind to think about them. These include:

• You can’t travel at the speed of light.
• Time is not fixed; it is relative.
• The speed of light is a universal constant—it is the same, independent of the motion of the source.

Einstein wrote:

“After ten years of reflection, such a principle resulted from a paradox upon which I had already hit at the age of sixteen: If I pursue a beam of light with the velocity c [the velocity of light in a vacuum], I should observe such a beam of light as a spatially oscillatory electromagnetic field at rest. However, there seems to be no such thing…”

- Autobiographical notes, 1949

I’ll try to explain a little as I understand it.

Our usual experience is that velocities are additive. Suppose I am on a moving train carriage and I throw a ball from the back of the carriage to the front. For an observer outside the train, that ball moves at the speed of the train plus the speed of the ball relative to me. But light behaves differently.

As you approach the speed of light, the energy required to keep accelerating approaches infinity. In effect, you can’t reach the speed of light. So an observer of a flying Einstein wouldn’t see light travelling from him to the mirror at twice the speed of light. What changes is time.

For the high-speed Einstein, the light would appear to travel away from him to the mirror and back at its usual immense speed. However, for an observer, what would only seem a moment for the high-speed Einstein might take years for the rest of us—the experience of time changes with velocity.

It’s a remarkable turn for a simple and fascinating question. It’s amazing to me that the young Einstein would both pose this question, continue work on it, and then think to question some of the most self-evident facts of our world as we experience it: that time is not fixed, that a speed cannot be reached, and of course, ultimately, that energy is matter.

The book Einstein’s Mirror is co-authored by my Dad (respect!). It’s full of photographs, fascinating stories, and the characters that moved physics forward. It includes the people, events and science central to another of Christopher Nolan’s films, Oppenheimer. Perhaps Christopher read it 🤔

Related Ideas to Einstein’s Mirror

Also see:

• Laplace’s Demon
• Redshift
• Looking back in time
• The Doppler Effect
• Sonic Boom
• The most beautiful equation

—

Earlier this year, we attended a showing of Christopher Nolan’s Interstellar at the Royal Albert Hall in London with Hans Zimmer’s soundtrack played by a live orchestra. It was a fantastic way to experience a remarkable film—a film that manages to make black holes, wormholes, and time slippage both understandable (largely) and part of the plot. It strikes me as an astonishing achievement for a mainstream film.

There's a lot of great FUI design to explore in Andor Season 2, by the talented folk at BLIND LTD, including:

• TIE Fighter cockpit displays

• targeting scopes

• industrial equipment

• medical devices

• tablets

• espionage tools

• desktop computers used throughout the Empire

Each planet introduces its own flavour of technology, but a consistent visual language helps tie everything together and makes it all feel unmistakably Star Wars.

TIE Fighter cockpit displays

The TIE Fighter cockpit displays are some of the most interesting interfaces in the series. I love how the main display is framed within a hexagonal screen and how the UI wraps around it. It creates such a unique look for the TIE Fighter.

The light blue UI colour looks great against darker backgrounds, although it struggles slightly when displayed over brighter scenes. It feels like a good candidate for an automatic light and dark mode. One of the most dramatic moments is when the entire display shifts to red.

I also love the triangular panel layouts. They may not be the most practical in the real world, but they immediately make the cockpit feel sci-fi and alien.

The controls feel daunting at first, which fits the moment where Andor gets into the ship for the first time. It looks unfamiliar, but then he realises it’s just like a motorcycle-style throttle, something familiar enough to figure out quickly.

View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize

Analogue devices

It’s good to see a few analogue devices in the series. There’s something great about seeing analogue devices reimagined in a futuristic setting. I'm sure the set department would have had a lot of fun with these!

View fullsize
View fullsize

Desktop devices

The design of these desktop devices is great. The screens feel very lo-fi, with unusual layouts that don’t try to mimic modern interfaces. The controls look a bit like an old NASA command centre aesthetic, using dials, knobs, and backlit LED buttons instead of keyboards (See Alien Romulus).

View fullsize
View fullsize
View fullsize
View fullsize
View fullsize

Viewfinders and scopes

The viewfinders have very interesting shapes, and the scrolling numbers feel reminiscent of old rangefinding equipment. The weapon scopes feel slightly more advanced and modern in comparison, which creates a nice contrast between an everyday item and a specialised weapon.

View fullsize
View fullsize
View fullsize
View fullsize

Elevator UI

An elevator is something that feels universal, it goes up and down, and the scrolling panel immediately communicates that motion.

What makes this interface work is how it suggests a familiar function, then adds details that make it feel different. The numbers are split into three separate slots, and the buttons seem to be mapped to those slots, making the display feel both familiar and alien at the same time.

Rhydonium fuel device

This display feels much more utilitarian, like something you’d expect to see on industrial equipment. It’s clearly designed for people who know what they’re doing, with an interface that assumes a level of expertise rather than explaining itself.

View fullsize
View fullsize
View fullsize
View fullsize

Handheld tablets

The tablet designs are very interesting. They use a format that already feels naturally useful, but push it in a different direction. It feels like what a tablet might have looked like if it had been designed in the 80s or 90s.

The interface feels restricted by the casing itself, with panel layouts dictated by the physical form rather than a clean digital grid. The buttons are basic, and the tablet itself is really chunky, and a bit cumbersome. But that’s part of what makes them great.

View fullsize
View fullsize
View fullsize
View fullsize

Empire OS

Here’s a nice glimpse into the operating system used within Empire headquarters. It feels clean, minimal, and almost clinical in its presentation. Everything is neat and tightly controlled.

The colour palette is very restrained, mostly black and white with touches of red, clearly aligned with the Empire’s visual language. It feels intentional, almost sterile, which reinforces the sense of order and control.

View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize

Various espionage tools

The espionage tools are really interesting, almost custom-made devices that don’t sit neatly in any single category. Each one feels like it’s built for a very specific function.

The code-breaking tool in particular has a strong presence. It feels foreign at first, but you could probably guess what it does just from how it behaves. Its rotating elements makes you think of combination locks, and the light turns green once it’s complete.

I also like the signal detecting tool which activates with what looks a bit like the HUDS+GUIS logo :)

View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize
View fullsize

Med-bay UI

The medbay interfaces are brutally minimalist, showing almost only data. They strip everything back to the essentials, with very little visual noise.

View fullsize
View fullsize
View fullsize

Google ferme les dernières portes. À partir de fin juin, avec Chrome 150 ou 151, les combines qui permettaient encore de garder uBlock Origin en vie dans le navigateur vont sauter une à une, et avec elles à peu près tous les bloqueurs de publicité qui reposaient sur l'ancienne plomberie de Chrome.

Petit retour en arrière. Une extension obéit à un cadre technique imposé par Google, le Manifest. Dans sa version 2, dite MV2, ce cadre laissait l'extension inspecter et modifier en direct chaque requête qu'une page web envoyait, ce qui donnait à un outil comme uBlock Origin, le bloqueur open source de Raymond Hill, sa fameuse capacité à filtrer dans le moindre recoin et à corriger ses règles en temps réel.

La version 3 change tout. Avec MV3, c'est désormais le navigateur qui bloque, à partir d'une liste de règles figée et plafonnée que l'extension se contente de lui fournir. Fini l'interception à la volée.

Google parle d'enterrer MV2 depuis 2019. Sauf que des soupapes traînaient encore. Une politique d'entreprise au nom à coucher dehors, ExtensionManifestV2Availability, plus quelques drapeaux internes et même une bidouille du registre Windows, permettaient de rallumer manuellement une extension que Chrome venait pourtant de couper.

Et c'est exactement ce bouquet de rustines qui disparaît. Le drapeau ExtensionManifestV2Disabled est déjà parti avec Chromium 150. La 151 doit emporter le reste, dont la fameuse politique et l'option AllowLegacyMV2Extensions, celles sur lesquelles s'appuyaient les derniers contournements. La bidouille du registre, elle, rendra l'âme dans la foulée.

Résultat, beaucoup d'utilisateurs ne retrouveront tout bonnement plus uBlock Origin dans leur liste d'extensions, sans le moindre bouton pour le ranimer.

Reste l'alternative maison, uBlock Origin Lite, signée du même auteur et compatible MV3. Sauf qu'allégée, ici, ce n'est pas un slogan : cette mouture refile au navigateur une liste de règles préchargée et perd au passage le filtrage dynamique, le nettoyage cosmétique poussé des pages et la possibilité de pointer soi-même l'élément gênant à virer. Elle stoppe les pubs classiques, beaucoup moins les formats modernes qui se réinventent en permanence.

Et Chrome n'est pas seul concerné. Edge et Opera, bâtis sur le même moteur Chromium, vont suivre la même pente et appliquer ces restrictions à leur tour. Firefox, de son côté, continue de faire tourner la mécanique MV2 et donc le uBlock Origin complet, alors que Brave et Vivaldi promettent de la maintenir sans s'engager pour autant sur le très long terme.

Officiellement, Google invoque la sécurité et la performance, l'idée étant qu'une extension capable d'éplucher tout votre trafic peut aussi jouer les mouchards. L'argument se tient. Il arrange juste drôlement bien une boîte dont l'essentiel des revenus vient précisément de cette publicité qu'on essaie de bloquer.

Si vous tenez à votre vieux bloqueur, passez sur Firefox.

Source : Bleeping computer

Qualcomm seems to be teasing its next-gen Snapdragon XR chipset targeted at standalone headsets, something the company says we’ll learn more about “soon.”

It’s uncertain whether Qualcomm is getting ready to announce its next-gen Snapdragon XR3 platform, or Gen 3 of its previously released XR2 chipset. What is certain though is XR’s biggest chip manufacturer has something new in store, teasing a “new reality” in a short clip posted on X.

Imagine #XR that's smarter and more immersive than anything you've experienced before. Something new is coming pic.twitter.com/2Of6E0BS2R

— Snapdragon (@Snapdragon) June 9, 2026

And while Qualcomm says we’re due to learn about it “soon,” there’s really no telling when that could be, as the company really hasn’t stuck to a set release schedule for its various XR chip announcements.

Notably, Qualcomm announced Snapdragon XR1 at AWE 2018 in May, Snapdragon XR2 at Snapdragon Tech Summit in December 2019, Snapdragon XR2 Gen 2 right before the launch of the Quest 3 in 2023, and Snapdragon XR2+ Gen 2 just before CES 2024.

There are however a few venues the company could announce its next iteration of Snapdragon XR. Qualcomm’s June 24th Investor Relations Day may deliver our first bit of insight into its first slate of hardware partners. There’s also the company’s big Snapdragon Summit in late September, which is notably during the exact same timeframe as Meta Connect.

Thus far, Meta has only teased what appears to be a new slate of smart glasses though, which in the past have integrated Qualcomm’s Snapdragon AR1 chipset. Meta is also seemingly preparing a puck-tethered thin and light standalone, codenamed ‘Phoenix’, although that’s reportedly been delayed to 2027.

Possibly a more likely (and timely) candidate is Pico’s upcoming Project Swan, which the Byte Dance-owned company teased in March to contain a separate flagship SoC with “more than 2× CPU and GPU performance vs XR2 Gen 2.” That’s supposed to launch globally sometime in late 2026, so we may learn a lot more fairly soon.

The post Qualcomm Teases Next-Gen Snapdragon XR Chipset, Possibly Debuting in Pico’s Next Flagship appeared first on Road to VR.