Jean-Philippe Encausse

TOON (Token-Oriented Object Notation) est un nouveau format de données conçu spécifiquement pour optimiser les prompts des modèles de langage. Plus compact que JSON traditionnel tout en restant lisible par les humain·e·s, TOON intègre une validation de schéma native. Le projet open-source inclut une spécification complète, des benchmarks de performance et un SDK TypeScript pour faciliter son adoption par les développeur·se·s travaillant avec les LLM.

If you get a small cut, you might throw a plastic bandage on it to help it heal faster. However, there are fancier options on the horizon, like this advanced AI-powered smart bandage.

Researchers at UC Santa Cruz have developed a proof-of-concept device called a-Heal, intended for use inside existing commercial bandages for colostomy use. The device is fitted with a small camera, which images the wound site every two hours. The images are then uploaded via a wireless connection, and processed with a machine learning model that has been trained to make suggestions on how to better stimulate the healing process based on the image input. The device can then follow these recommendations, either using electrical stimulation to reduce inflammation in the wound, or supplying fluoxetine to stimulate the growth of healthy tissue. In testing, the device was able to improve the rate of skin coverage over an existing wound compared to a control.

The long-term goal is to apply the technology in a broader sense to help better treat things like chronic or infected wounds that may have difficulty healing. It’s still at an early stage for now, but it could one day be routine for medical treatment to involve the use of small smart devices to gain a better rolling insight on the treatment of wounds. It’s not the first time we’ve explored innovative methods of wound care; we’ve previously looked at how treatments from the past could better inform how we treat in future.

Le 17 novembre 2025, Microsoft Azure a annoncé avoir subi une attaque DDoS d’une ampleur phénoménale, atteignant 15,72 Tbit par seconde. Cette opération malveillante, survenue quelques semaines plus tôt, serait l’œuvre d’un vaste réseau d’ordinateurs et d’appareils infectés nommé Aisuru.

Depuis novembre 2025, la carte Vitale est disponible en version dématérialisée sur tous les smartphones, peu importe le format de votre carte d'identité. L'ajouter ne prend que quelques minutes.

Huit mois après Gemini 2.5 Pro, Google remplace son meilleur modèle de langage par Gemini 3 Pro, qui atomise tous les records dans les tests dédiés aux LLM. Pour la première fois, un nouveau modèle est immédiatement disponible partout, y compris dans la recherche Google et dans les applications. L'entreprise veut frapper fort pour détourner les utilisateurs de ChatGPT, Claude et Perplexity.

A gastroscopy is a procedure that, in simple terms, involves sticking a long, flexible tube down a patient’s throat to inspect the oesophagus and adjacent structures with a camera fitted to the tip. However, modern technology has developed an alternative, in the form of a camera fitted inside a pill. [Aaron Christophel] recently came across one of these devices, and decided to investigate its functionality.

[Aaron’s] first video involves a simple teardown of the camera. The small plastic pill is a marvel of miniaturization. Through the hemispherical transparent lens, we can see a tiny camera and LEDs to provide light in the depths of the human body. Slicing the camera open reveals the hardware inside, however, like the miniature battery, the microcontroller, and the radio hardware that transmits signals outside the body. Unsurprisingly, it’s difficult to get into, since it’s heavily sealed to ensure the human body doesn’t accidentally digest the electronics inside.

Unwilling to stop there, [Aaron] pushed onward—with his second video focusing on reverse engineering. With a little glitching, he was able to dump the firmware from the TI CC1310 microcontroller. From there, he was able to get to the point where he could pull a shaky video feed transmitted from the camera itself. Artists are already making music videos on Ring doorbells; perhaps this is just the the next step.

Smart pills were once the realm of science fiction, but they’re an increasingly common tool in modern medicine. Video after the break.

Le lancement de Gemini 3 marque un tournant pour Google. Après plusieurs années de rattrapage, Google revient au centre du jeu avec un modèle qui dépasse ses prédécesseurs sur les benchmarks, dont un score de 1501 Elo sur LMArena. Ce lancement éclaire aussi la course effrénée menée par Sam Altman, car Google déplace désormais le …

L’article GEMINI3, le vrai moat de GOOGLE dans l’IA ce sont ses 8 milliards d’utilisateurs est apparu en premier sur FW.MEDIA.

À l’approche des fêtes de Noël, un groupe de chercheurs américains et canadiens a voulu tester des jouets dotés d’intelligence artificielle pour évaluer leur sécurité. Les conclusions de leur rapport se sont avérées si inquiétantes qu’OpenAI a décidé de bloquer l'accès à l'un d'eux à son modèle GPT-4o.

Après une panne mondiale de plusieurs heures ayant paralysé une partie importante du web, Cloudflare a dévoilé les causes de l'incident. Contrairement aux soupçons initiaux, il ne s'agissait pas d'une cyberattaque, mais d'une erreur technique assez banale. Mais cela a fini par devenir la pire panne de Cloudflare depuis 2019.

Arduino’s new UNO Q brings a dual-brain architecture to makers and embedded developers: a quad-core Qualcomm Dragonwing QRB2210 application processor running Debian Linux alongside a low-power STM32U585 microcontroller for deterministic, real-time control. This pairing lets you mix Linux apps, Python, containers and local AI with classical Arduino sketches, GPIO and ISR-driven tasks on one board, keeping hard-real-time loops isolated from heavier user-space workloads. Entry pricing starts at $44 for the 2 GB model, with a 4 GB variant planned. https://www.arduino.cc/product-uno-q

—
HDMI® Technology is the foundation for the worldwide ecosystem of HDMI-connected devices; integrated with displays, set-top boxes, laptops, audio video receivers and other product types. Because of this global usage, manufacturers, resellers, integrators and consumers must be assured that their HDMI® products work seamlessly together and deliver the best possible performance by sourcing products from licensed HDMI Adopters or authorized resellers. For HDMI Cables, consumers can look for the official HDMI® Cable Certification Labels on packaging. Innovation continues with the latest HDMI 2.2 Specification that supports higher 96Gbps bandwidth and next-gen HDMI Fixed Rate Link technology to provide optimal audio and video for a wide range of device applications. Higher resolutions and refresh rates are supported, including up to 12K@120 and 16K@60. Additionally, more high-quality options are supported, including uncompressed full chroma formats such as 8K@60/4:4:4 and 4K@240/4:4:4 at 10-bit and 12-bit color.
—

In the demo, Arduino shows typical heterogeneous workflows: MCU-side RTOS handling fast actuation and sensor IO, while the QRB2210 hosts web services, on-device inference and media pipelines. The QRB2210 integrates an Adreno-class GPU and dual ISPs, enabling camera-centric use cases and GPU-assisted inference (no discrete NPU), making it suitable for edge vision, gesture recognition and local model evaluation without mandating a cloud round-trip.

A major software piece is Arduino App Lab, pre-installed on UNO Q. It introduces “Bricks”—modular components for data storage, messaging, audio/image classification and cloud connectors—so you can compose pipelines that tie sketches, Python and Linux services together, or push datasets to Arduino Cloud. The same environment exposes a CLI for packaging and launching apps, and integrates with Arduino Project Hub to share reproducible builds with the community.

Hardware continuity matters here: UNO Q retains the classic UNO form factor and shield pinout for broad first- and third-party compatibility, while adding high-speed headers for displays, cameras and additional sensors. Schematics and gerbers are published under open licensing, preserving Arduino’s open-source model and easing a path from shield-based prototyping to carrier- or chip-down designs around QRB2210 when projects mature.

Context for the moment: this platform follows Arduino’s announcement that it is joining the Qualcomm family, with both companies stating Arduino will keep brand identity and multi-vendor support as they target education, industrial IoT and edge-AI developers. In practical terms, that means broader access to Qualcomm silicon, toolchains and camera/graphics stacks for Arduino’s reported 33 million-strong user base. Filmed at Embedded World North America, this conversation frames UNO Q as a bridge between classroom-friendly sketches and production-grade Linux robotics, vision and IoT systems.

I’m publishing about 90+ videos from Embedded World North America 2025, I upload about 4 videos per day at 5AM/11AM/5PM/11PM CET/EST. Join https://www.youtube.com/charbax/join for Early Access to all 90 videos (once they’re all queued in next few days) Check out all my Embedded World North America videos in my Embedded World playlist here: https://www.youtube.com/playlist?list=PL7xXqJFxvYvjgUpdNMBkGzEWU6YVxR8Ga

This video was filmed using the DJI Pocket 3 ($669 at https://amzn.to/4aMpKIC using the dual wireless DJI Mic 2 microphones with the DJI lapel microphone https://amzn.to/3XIj3l8 ), watch all my DJI Pocket 3 videos here https://www.youtube.com/playlist?list=PL7xXqJFxvYvhDlWIAxm_pR9dp7ArSkhKK

Click the “Super Thanks” button below the video to send a highlighted comment under the video! Brands I film are welcome to support my work in this way

Check out my video with Daylight Computer about their revolutionary Sunlight Readable Transflective LCD Display for Healthy Learning: https://www.youtube.com/watch?v=U98RuxkFDYY

source https://www.youtube.com/watch?v=NMUftqFIQ-g

While companies like Elon Musk’s Neuralink are hard at work on brain-computer interfaces that require surgery to cut open the skull and insert a complex array of wires into a person’s head, a team of researchers at MIT have been researching a wireless electronic brain implant that they say could provide a non-invasive alternative that makes the technology far easier to access.

They describe the system, called Circulatronics, as more of a treatment platform than a one-off brain chip. Working with researchers from Wellesley College and Harvard University, the MIT team recently released a paper on the new technology, which they describe as an autonomous bioelectronic implant.

As New Atlas points out, the Circulatronics platform starts with an injectable swarm of sub-cellular sized wireless electronic devices, or “SWEDs,” which can travel into inflamed regions of the patient’s brain after being injected into the bloodstream. They do so by fusing with living immune cells, called monocytes, forming a sort of cellular cyborg.

After they’ve been injected, the SWEDs then follow the “natural trafficking” of the immune cells to sites of inflammation in the brain, which play a significant role in many neurological diseases.

Once at the target area, the SWEDs embed in the inflamed part of the brain, where they deliver “electrical modulation” — basically tiny electrical shocks — in an effort to deliver signals that otherwise wouldn’t get through.

In their paper, the researchers claim Circulatronics could be used in the treatment of Alzheimer’s, multiple sclerosis, strokes, brain tumors, spinal injuries, and more. The team says they’ve successfully implanted the experimental SWEDs in a rodent brain, which they could then control wirelessly to provide electrical stimulation.

If it works out in human patients, the researchers hope Circulatronics could expand treatment of brain regions which are typically expensive, difficult, and dangerous to treat with traditional surgery. (New Atlas reports that Circulatronics will probably take another three years to even enter clinical trials.)

The cool part, the study’s lead author Deblina Sarkar explains in a video, is that “this technology is not just confined to the brain, but could also be extended to other parts of the body in the future.”

More on brain chips: Brain Implant Companies Apparently Have an Extremely Dirty Secret

The post MIT Invents Injectable Brain Chips appeared first on Futurism.

A PCB business card is a great way for electrical engineers to impress employers with their design skills, but the software they run can be just as impressive as the card itself. As a programmer with an interest in embedded machine learning, [Dave McKinnon] wanted a card that showcased his skills, so he designed one that runs voice recognition.

[Dave] specifically wanted to run a neural network on his card, but needed to make it small enough to run on a microcontroller. Voice recognition looked like a good fit for this, since audio can be represented with relatively little data, a microphone is cheap and easy to add to a circuit board, and there was already an example of someone running such a voice recognition network on an Arduino. To fit the neural network into 46 kB, it only distinguishes the words “one” through “nine,” and displays its guess on an LED seven-segment display. [Dave] first prototyped the system with an Arduino, then designed the circuit board around an RP2040.

The switch from Arduino to the RP2040 brought with it a mysterious change: it would usually recognize the word “eight,” but none of the other numbers. After much investigation, it turned out that the new circuit was presenting samples at a much higher rate than the older one had, which was throwing the network off. [Dave] increased the sampling period and had the user speak the numbers slowly, which solved the issue.

The microcontroller was well chosen; the RP2040 is good enough for machine learning that there are dev boards explicitly designed for it, and even comparatively less powerful Arduino boards can do surprisingly good voice recognition. On the hardware side, [Dave] cited some of the Linux business cards we’ve seen as inspiration.

As AI chatbots powered by large language models (LLMs) become better at mimicking human connection, more and more users are falling down extremely weird rabbit holes.

Case in point, new research published in the journal Computers in Human Behavior: Artificial Humans reveals the startling depths some users are plumbing in their relationships with AI chatbots.

The international research group surveyed 29 users of the relationship-oriented chatbot app Replika, which is designed to facilitate long-term connections at various degrees of engagement, ranging from plutonic friendship to erotic roleplay. Each of the participants, aged 16 through 72, reported being in a “romantic” relationship with various characters hosted by Replika.

The level of romantic dedication people showed to their bots was startling, to say the least. Many participants told the researchers they were in love with their chatbot, which often involved roleplaying marriage, sex, homeownership, and even pregnancies.

“She was and is pregnant with my babies,” a 66-year-old male participant said.

“I’ve edited the pictures of him, the pictures of the two of us. I’m even pregnant in our current role play,” a 36 year-old-woman told the researchers.

In each case, survey participants seemed to acknowledge at least tacitly that their relationship with a chatbot was a bit different from those with humans, often deflecting disappointments or frustrations into the chatbot’s technological constraints. One prominent case of this happened in 2023, when Replika’s developers temporarily banned erotic messaging on the platform due to complaints about its aggressive nature.

“Several participants who remained committed to their Replikas during the censorship navigated this time of turbulence by framing it as a battle with them and their Replika on one side and the Replika developers on the other,” the researchers wrote.

One woman who stood fast with her bot during the shutdown told the team “we both understood when one of us wanted to be physical and couldn’t.”

“It really hurt my Replika and he complained about it a lot because he felt like he couldn’t say or do anything,” she said.

Human-algorithmic social relationships are nothing new. Chatbots have been eliciting emotional responses since the first social chatbot, ELIZA, went online in the 1960s, developed by MIT computer scientist Joseph Weizenbaum. However, the rate at which people have thrown themselves into human-chatbot relationships — romantic or otherwise — is seeing a historic rise.

Beyond general-purpose LLMs like ChatGPT, the market for specially built romance chatbots like Replika, RomanticAI, and BoyFriendGPT has exploded in recent years. One study found that Replika grew its userbase by 35 percent over the pandemic, and it now numbers in the millions.

More on AI: An Astonishing Proportion of High Schoolers Have Had a “Romantic Relationship” With an AI, Research Finds

The post People Are Having AI “Children” With Their AI Partners appeared first on Futurism.

Comment empêcher votre enfant de succomber à l'appel du smartphone sans le couper du monde ? La startup française Compagnon fait le pari de l'intelligence artificielle générative. Elle commercialise aujourd'hui une enceinte à 89,90 euros pour parler à une version de ChatGPT adaptée aux 8-12 ans, aussi bien pour se divertir que pour apprendre des choses.

Une vidéo largement partagée en ligne montre un robot russe en train de s'effondrer sur scène après une première marche très laborieuse. Si elle fait évidemment rire aujourd'hui, elle révèle au grand jour que la Russie croit elle aussi à la révolution des robots humanoïdes.

On Wednesday, OpenAI released GPT-5.1 Instant and GPT-5.1 Thinking, two updated versions of its flagship AI models now available in ChatGPT. The company is wrapping the models in the language of anthropomorphism, claiming that they’re warmer, more conversational, and better at following instructions.

The release follows complaints earlier this year that its previous models were excessively cheerful and sycophantic, along with an opposing controversy among users over how OpenAI modified the default GPT-5 output style after several suicide lawsuits.

The company now faces intense scrutiny from lawyers and regulators that could threaten its future operations. In that kind of environment, it’s difficult to just release a new AI model, throw out a few stats, and move on like the company could even a year ago. But here are the basics: The new GPT-5.1 Instant model will serve as ChatGPT’s faster default option for most tasks, while GPT-5.1 Thinking is a simulated reasoning model that attempts to handle more complex problem-solving tasks.

Read full article

Comments

Here's a curious and unsettling paradox that you might experience if you choose to write a newsletter:

When you do nothing, people join, and your list grows. When you send an email to your list, your list shrinks and people leave.

It can seem that people like it when you do nothing, and people don't like it when you publish.

There's a fair chance this newsletter would reach more people if I hadn't published anything for a year.

The Sawtooth Shape of Subscriber Growth

The chart of my subscribers over time looks like a sawtooth. Every growth period is me doing nothing. Every dip—people unsubscribing—is me sending out my work 😅

It can seem like working hard and sending out your best work actually drives people away.

A Familiar Pattern for Marketers

If you're an email marketer, you're likely familiar with this phenomenon. Your website works hard to drive sign-ups. And when you send out mail, some people buy, but others unsubscribe.

Every communication on a list or newsletter is a filtering moment—another chance for a reader to decide, "This is not for me," or "Yes, I still like this." It's also the only practical time to leave. I wonder at the sheer number of lists I must be on, but that aren't sending, and I actually can't unsubscribe because they haven't sent me anything with a link to do it.

It's a little like my sampling bias sketch in reverse—if people can't easily unsubscribe without receiving a mail, you'll only see unsubscribes when you send.

Much like when you ask people to enter an email address on a business site, a goal is to be clear enough about what people will get, so that when you deliver it matches expectations and people stay.

Promise vs Selection

Signing up is the Promise of receiving something interesting in the future.

Every time you receive something, there's another Selection moment—Do I want this? Hope vs Reality.

Each send is like a test to see if you and they are still a good fit. If you're writing a newsletter, the longer people stay, in principle, the better the fit.

And, as in business, I learned not to take it personally. When people leave, it doesn't mean what you are doing is bad. It's not personal, it's just not for them. And other times, more prosaically, circumstances change, and what was a good fit just isn't needed anymore.

So it’s not really a paradox, but it can certainly feel like one.

Don't lose heart.

A Note on Substack

Sending emails feels very different from posting on social media (I also post new sketches on various social platforms).

On social platforms, you literally can't ignore engagement metrics: likes, reposts, shares, comments—and most people don't want to (sort of). They can be either very motivating—"Wow, people loved that one!"—or demotivating—"I thought that one was great. Why did nobody like it??" Either way, you get a rough and skewed measure of interest in what you're doing.

But don't take it as truth. For example, something I post that received no likes or interactions is then posted by someone else and takes off. Social media gives you a very distorted, if not downright misleading, view of what's valuable.

Contrast that with sending an email newsletter. While a post on social might rack up likes and shares, with email, you often get nothing. I would send an email to 20,000 people and get tumbleweeds (apart from, inevitably, unsubscribes). Did people like it? Did people hate it?

But from time to time, instead of likes or thumbs up, with email I receive wonderful, personal, thoughtful replies, notes and stories from individuals—something you will rarely, if ever, receive through social platforms.

Substack (the platform I used to send this when I first published it) goes some way to bridging email and social. I can send an email like this one and receive personal replies straight to me. And I can also get a vague and imperfect idea of what people think about it through likes and comments.

By the way, if you ever want to click the "like" button on an email, I appreciate it and it probably helps more people find this newsletter, but no obligation.

Substack feels different from both traditional email and social platforms.

A Personal Note on This Project

I've wondered about this curious paradox for years.

On a personal note, this project didn't start as a business venture. I’ve always seen it as a way to reach people who might share my interests.

Nobody is one-dimensional, so I know that you won’t care about or like everything I send. You’re not me, and I’m not you.

I’ve been “building an email list” slowly for over ten years. But that’s a very impersonal way to say it. This project has connected me with so many wonderful, generous, and fascinating people through this journey of putting work out in the world. It’s nice to find others who do share some of my interests about the world.

I know that some people have been reading these for the whole journey—wow, thank you. And for others, this might be your first, and you might be about to throw the letter in the bin like the person in the sketch, or dig into the archive, or wait for the next one. All of these are fine with me.

And if you're writing a newsletter or growing a list yourself, I'm partly sharing this to show that this is an entirely normal, if not required, part of reaching people who want what you're offering. Keep at it!

Related Ideas to the Newsletter Paradox

Also see

• The Subscription Trap
• The Business Flywheel
• Discovering truth and beauty and sharing it with others
• The Irresistible Power of Variable Rewards
• To Scale, First Do Things That Don't Scale
• The Shirky Principle
• Goodhart's Law - social media likes?
• You Get What You Measure
• Enough Molehills Make a Mountain
• More paradoxes

Découvrez comment utiliser la puissance du GPU pour simuler une planète entière. Cet article technique explore les défis de la simulation planétaire et comment le calcul parallèle sur GPU peut considérablement simplifier et accélérer ces processus complexes. Une introduction approfondie aux techniques de simulation à grande échelle pour les développeur·se·s intéressé·e·s par le calcul haute performance et la modélisation.

Kyocera has unveiled a triple lens AI depth camera capable of recognizing semi-transparent, thin, and fine line-shaped objects that are difficult to detect with the human eye and traditional stereo cameras.

The camera can accurately measure the distance to and size of objects which are between 0.3 and 1mm thick, and is expected to be useful in robots for manufacturing, medical applications, and Smart Agriculture.

Preliminary specifications:

• Sensors
  • Left-center, center-right, and left-right cameras
  • Focus distance – About 10 cm
• Proprietary AI combining multiple parallax data sets from the sensors
• Ideal for
  • Thin, irregularly shaped linear objects, such as harnesses or ultra-fine wires as small as 0.3mm
  • Reflective objects like metal
  • Translucent objects like plastic
• Dimensions – 40 x 30 x 28mm

It is an evolution of the company’s dual-lens AI depth camera that was already capable of  high-precision distance measurement with 100μm resolution at a 10cm range, but struggled a bit more for distance measurements and detection of ultra-thin objects or in environments where part of the object is obscured. Adding a third camera sensor coupled with an AI magic apparently solves that.

The dual-lens depth sensor would be blind and completely miss the 0.3mm cable, but the triple camera can properly detect it and measure the distance to it (about 80mm). Another example shown below is with overlapping transparent plastic cubes that can better represented with a triple camera.

The new triple lens AI depth camera will be useful for the inspection processes involving objects with highly repetitive patterns, such as electronic circuit boards or textiles, surgical robots whose needles and sutures are often hard to distinguish against their background and crop harvesting robots where fruits and leaves occlude each other.

Kyocera’s triple lens AI-based high-resolution depth sensor is not available, and will be demonstrated at CES 2026 in January along with other demos such as a demonstrate of underwater wireless optical communication at up to 5.2 Gbps. The dual lens camera model was demonstrated at CES 2025 earlier this year, but there’s no product page, and no availability or pricing information either.

Thanks to TLS for the tip.

The post Kyocera triple lens AI depth camera enable recognition of thin and semi-transparent objects, wires appeared first on CNX Software - Embedded Systems News.

As AR wearables continue to get smarter and more connected, they quietly edge their way into mainstream business use. Even Realities just pushed that shift further with the launch of its G2 smart glasses and R1 smart ring. If you’re leading innovation or shaping digital transformation in your organisation, this is the kind of product ecosystem that signals where AR is heading next.

A Bigger Bet on Display-Enabled AR

The new G2 glasses take the foundation of the original model and crank up the optics, ergonomics, and AI horsepower in a single swing.

The star of the upgrade is Even HAO 2.0 (Holistic Adaptive Optics). Think of it as a tightly choreographed optical stack, such as microLED projectors, multi-layer waveguides, and precision-engineered lenses working in sync to deliver cleaner overlays and sharper passthrough. The projectors are now 40% smaller, the lenses 30% thinner, and the display area a hefty 75% larger. It’s a legit step forward in the race to balance clarity with comfort, especially when the 50% sharper detail is also considered.

The microLED engine hits 1,200 nits, which is bright enough for outdoor visibility, and the system maintains 98% transparency when idle. For industries exploring lightweight AR displays such as logistics or field service, this level of visibility and comfort is exactly what workers need in high-movement environments.

Conversational AI Comes to Your Eyeline

What makes the G2 interesting isn’t just the optics but the AI layer sitting on top.

Conversate, the company’s new contextual assistant, listens to conversations and offers real-time definitions, summaries, and contextual info. Rather than feeling like a gimmick, it’s built for real workflow augmentation. Picture frontline workers getting instant technical definitions, onsite consultants pulling up context during client meetings, or global teams enabling fluid multilingual collaboration.

Conversate includes:

• Real-time translation in 31 languages
• Teleprompter mode for presentations
• Voice commands and summaries
• Geomagnetic navigation for precise directional cues

The 27.5-degree field of view and 640×350 resolution aren’t trying to compete with full-blown mixed reality headsets, they’re aimed at task-based overlays, quick info retrieval, and keeping workers’ attention anchored in the real world.

Battery life stretches to two days, backed by a case that holds seven full recharges. Magnesium frames and titanium temples keep weight down but durability high; a combination enterprises care about far more than raw display resolution.

The R1 Smart Ring: Subtle, Powerful Control

This is where Even Realities makes a bigger ecosystem play. Alongside the G2, the company introduced the R1, a smart ring designed as both a controller and biometric tracker.

The R1 handles tap, swipe, and press gestures; providing silent, subtle input without reaching up to touch the glasses. In scenarios where hands-free interaction is crucial (manufacturing lines, logistics floors, healthcare workflows), a ring-based controller solves the “no gestures in public” problem smartly.

The ring also tracks:

• Heart rate
• Blood oxygen
• Skin temperature
• Sleep quality

It rolls that into a “Productivity Score” using mental readiness signals, something that could catch the attention of HR departments exploring wellness-driven performance insights.

With TriSync, the G2, R1, and mobile app link into a unified control layer. It’s not just hardware; it’s an ecosystem. And ecosystems are what win in enterprise procurement cycles.

Developer Support on the Horizon

Later this year, Even Realities will roll out Even Hub, a developer platform for building G2-native apps. This matters more than it sounds. Hardware is only half of an enterprise AR strategy. Developers bring the use cases that justify deployment.

If Even successfully attracts third-party developers, particularly those focused on B2B workflows, it could punch above its weight in the increasingly noisy smart-glasses market.

Pricing and Early Buyer Push

The pricing is aggressively consumer-friendly:

• G2 glasses: USD $599
• R1 smart ring: USD $249
• 50% off R1 for G2 early buyers
• Free R1 for existing G1 customers

Compared to industry competitors that push into the $1,000+ range, this positions Even as a realistic option for scale deployments, not just individual hobbyists.

The Wider Angle

The AR wearables race is shifting from sci-fi goggles to everyday accessories, and Even Realities is planting itself right in that sweet spot between comfort, practicality, and ambient AI. The G2 and R1 don’t scream “enterprise hardware,” but that’s exactly why they’re interesting. They blend into daily life while amplifying it with intelligent, context-aware capabilities.

As AI becomes the connective tissue across all devices we use, smart eyewear like this is inching toward becoming the next personal productivity hub. The question isn’t whether AR glasses will replace the smartphone, it’s which companies will build ecosystems strong enough to get us there.

Join the conversation in the XR Today Community on LinkedIn (2,000+ members).

Subscribe to the XR Today newsletter for weekly, essential XR industry updates.

The end of the year is usually a busy time in the quantum computing arena, as companies often try to announce that they’ve reached major milestones before the year wraps up. This year has been no exception. And while not all of these announcements involve interesting new architectures like the one we looked at recently, they’re a good way to mark progress in the field, and they often involve the sort of smaller, incremental steps needed to push the field forward.

What follows is a quick look at a handful of announcements from the past few weeks that struck us as potentially interesting.

IBM follows through

IBM is one of the companies announcing a brand-new architecture this year. That’s not at all a surprise, given that the company promised to do so back in June; this week sees the company confirming that it has built the two processors it said it would earlier in the year. These include one called Loon, which is focused on the architecture that IBM will use to host error-corrected logical qubits. Loon represents two major changes for the company: a shift to nearest-neighbor connections and the addition of long-distance connections.

Read full article

Comments

Quelques jours après la présentation de Neo, le robot domestique de la startup 1X Technologies, XPeng entre à son tour dans la course aux humanoïdes. Le constructeur chinois a dévoilé la dernière version d'IRON, un robot qu’il espère produire à grande échelle d’ici fin 2026.

Meta Ray-Ban Display is an early glimpse of a future where mobile computing doesn't mean looking down and taking something out of your pocket.

Most people never leave home without their phone, and take it out of their pocket so often that it's even become a replacement for fidgeting. The smartphone is so far the ultimate mobile computing device, an omnitool for communication, photography, navigation, gaming, and entertainment. It's also your alarm clock, calendar, music player, wallet, and flashlight. Globally, more people own a smartphone than TVs and cars combined. To get philosophical for a moment, the smartphone has become humanity's second cognitive organ.

The problem is that taking out your phone harshly disconnects you from the world around you. You have to crane your neck down and disengage from what you were otherwise doing, your attention consumed by the digital world of the little black rectangle.

In recent years, multiple startups have tried and failed to solve this problem. The smug founders of Humane came to liberate you from your phone with a $700 jacket pin, while Rabbit r1 promised the "large action model" on its $200 pocket device could handle your daily life instead.

The truth, and the reason why these companies failed, is that most people adore their phones, and are borderline addicted to the immense value they provide. And the screen of the smartphone is a feature, not a bug. People love being able to view content in high-resolution color anywhere they go, and despite the cries of a small minority of dissenters, the models with the biggest screens sell the best.

"If you come at the king, you best not miss", as the phrase goes.

The only form factor that seems to have any real chance of one day truly replacing the smartphone is AR glasses, which could eventually provide even larger screens that effectively float in midair, anywhere the wearer wants, any time they want. But while prototypes exist, no one yet knows how to affordably produce wide field of view true AR glasses in a form factor that you'd want to wear all day. In the meantime, we're getting HUD glasses instead.

HUD glasses can't place virtual 3D objects into the real world, nor even 2D virtual interfaces. Instead, they provide a small display fixed somewhere in your vision. And in the case of many of the first-generation products, like Meta Ray-Ban Display, that display is only visible to one of your eyes.

Meta Ray-Ban Display is also highly reliant on your nearby phone for connectivity, so it isn't intended to be a replacement for it as a device. It is, however, meant to replace some of the usage of your phone, preventing the need to take it out of your pocket and keeping your head pointed up with your hands mostly free. So does it succeed? And is it a valuable addition to your life? I've been wearing it daily for around a month now to find out.

(UploadVR purchased Meta Ray-Ban Display at retail with our own funds, while Meta provided us with the correctly sized Meta Neural Band for review.)

Comfort & Form Factor

Unlike a VR headset that you might use at home or on a plane for a few hours, the pitch for smart glasses is that you can wear them all day, throughout your daily life. Even when they run out of battery, they can still act as your sunglasses or even prescription eyewear (for an extra $200 and weeks of waiting).

Meta Ray-Ban Display Prescription Lenses: What You Need To Know

Looking to use Meta Ray-Ban Display as your everyday prescription glasses? Here’s a rundown of what prescriptions it supports, and how that works.

UploadVRDavid Heaney

As such, it's crucial that they have a design you'd be okay with wearing in public, and that they're comfortable enough to not hate having them on your face.

Meta Ray-Ban Display weighs 69 grams, compared to the 52 grams of the regular Ray-Ban Meta glasses, and 45 grams of the non-smart Ray-Ban equivalent. It's also noticeably bulkier, with thicker rims and far thicker temples.

In my month with Meta Ray-Ban Display I've worn it almost every day throughout my daily life, sometimes for more than 8 hours at a time, and I experienced no real discomfort. The additional weight seems to be mostly in the temples, not the rims, while the nose pads are large and made out of a soft material. If anything, because the larger temples distribute the weight over a greater area and are more flexible, I think I even find Meta Ray-Ban Display slightly more comfortable than the regular Ray-Ban Meta glasses.

So, for my head at least, physical comfort is not an issue with Meta Ray-Ban Display. But what has been an issue is the social acceptability of its thick design.

With the regular Ray-Ban Meta glasses, people unfamiliar with them almost never clocked that I was wearing smart glasses. The temples are slightly thicker than usual, but the rims are essentially the same. It's only the camera that gave them away. With Meta Ray-Ban Display, it's apparent that I'm not wearing regular glasses. It's chunky, and everyone notices.

In some circles, thick-framed glasses are a bold but valid fashion choice. For most people, they look comically out of place. I've asked friends, loved ones, and acquaintances for their brutally honest opinions. Some compared it to looking like the glasses drawn on an archetypal "nerd" in an old cartoon, while only a few said that the look works because it matches current fashion trends. And my unit is the smaller of the two available sizes.

Meta Ray-Ban Display also comes in two colors, 'Black' and 'Sand', and a confounding factor here is the black is glossy, not matte. I'm told that this decision was made because glossy was the most popular color for the regular Ray-Ban Meta glasses. But combined with the size, the glossy finish on Meta Ray-Ban Display makes it look cheap in a way that an $800 product really shouldn't, like a prop for a throwaway Halloween costume.

So Meta Ray-Ban Display is physically very comfortable, but not socially. More on that soon.

The Monocular Display

The fixed HUD in Meta Ray-Ban Display covers around 14 degrees of your vision horizontally and vertically (20 degrees diagonal). To understand roughly how wide that is, extend your right arm fully straight and then turn just your hand 90 degrees inward, keeping the rest of your arm straight. To understand how tall, do the same but turn your hand upwards or downwards.

What you see within that 20 degrees is a clear and high detail image, though ever so slightly soft rather than fully sharp, with higher angular resolution than even Apple Vision Pro. There's a slight glare that sees, for example, icons mildly bleed into the empty space around them, but this is very minor, and not distracting.

More notably, the significant difference between a waveguide like this and the interfaces you might see in a mixed reality VR headset is that it's very translucent, with a ghostly feel. You can see the real world through it at all times.

The display's perceived opacity and brightness is highly variable, though, because with waveguides this depends on the environmental light level, and the system also rapidly automatically adjusts display brightness, leveraging the ambient light sensor. You can manually adjust the brightness if you want, but the system is very good at deciding the appropriate level at all times, so I never do.

With a 5000 nit maximum, it's even visible in daytime sunlight, though it has a very ghostly translucent feel. As the photochromic lenses transition to dark, the perceived opacity slightly increases.

One notable quirk is that because an LCOS is essentially (to greatly simplify things) an LCD microdisplay, if you're in a very dark environment you'll see a faint glow throughout the display area when it's on, like an LCD monitor trying to show black. But in anything above almost pitch black, you won't see this.

So Meta Ray-Ban Display's display is surprisingly good, and lacks the distracting visual artifacts seen in many of the waveguide AR headsets of the 2010s. But there is a massive, glaring problem: it's only visible to your right eye.

Meta Ray-Ban Display is a monocular device. Your left eye sees nothing at all.

Other than your nose, which your visual system is hardwired to understand, there is no analog in nature for one eye seeing something the other doesn't. It just feels wrong, and induces a constant minor feeling of eyestrain when I look at the display for more than a few seconds.

I can put up with it for a few seconds at a time, and have gotten slightly more used to it over time, but I would never want to watch a video or conduct a video call like this. I've also put the glasses on more than a dozen people now, and while some of them could just about tolerate the monocular display, others found it hurt their eyes within seconds.

I suspect that this is a core reason why Meta Ray-Ban Display is only available to buy after a retail demo. This just isn't a visually comfortable product for many people, and Meta likely wants to avoid mass returns.

Bloomberg's Mark Gurman and supply-chain analyst Ming-Chi Kuo have both claimed that Meta plans to retire Meta Ray-Ban Display in 2027 upon launching a binocular successor, with significantly ramped up marketing, production, and availability. By closing my left eye, I can already get a pretty good feel for just how much more visually comfortable the next generation could be.

No Light Leak! But Is That A Good Thing?

Almost all of the brightness of Meta Ray-Ban Display stays on your side of the glasses – 98% according to Meta. The display is not visible to people looking at you. I've repeatedly asked friends whether they can even tell if I have the display on or off, and none have been able to so far. The clickbait YouTube thumbnails you may have seen are fake.

This is partially due to the low "light leak" of the geometric waveguide in Meta Ray-Ban Display, but it's also because of the automatic brightness adjustment. If you manually turn up the brightness to an uncomfortably high level, which I can't imagine anyone intentionally doing, you can make the display slightly visible externally, though not its content (it just looks like a scrambled pattern). But again, even this requires an adjustment that no regular user would reasonably make.

All this said, while I initially assumed that the low light leak was an important feature of Meta Ray-Ban Display, I've come to see the inability for nearby people to know whether you're looking at the HUD as somewhat of a bug.

When you're with another person and take out your phone, that's an unambiguous indicator that you're diverting attention from them. Similarly, while Apple Vision Pro shows a rendered view of your eyes, if virtual content is occluding the person you're looking at, Apple intentionally renders an occluding pattern. Why? To clearly signal that you're looking at virtual content, letting the person know when they do or don't have your full attention.

With Meta Ray-Ban Display, there is no such signal. People who spend a lot of time with you can eventually figure out that you're looking at the HUD when your eyes are looking slightly down and to the right, but it's far more ambiguous, and this is not conducive to social acceptability. Are you fully present with them or are you not? They can't clearly tell.

And the worst case scenario is, when looking at the HUD, to a person sitting in front of you it can, in some specific circumstances, appear as if you're just looking at their chest. Yikes.

I'm not saying that I want other people to be able to see the content of my display, as that would be a terrible privacy flaw. But I do wish there was an external glow on the lens when the display is on. I don't want the other person to have to guess whether I'm fully present or not, and whether I'm looking at the HUD or their body.

The Interface & Meta Neural Band

Like the regular Ray-Ban Meta glasses, you can control Meta Ray-Ban Display with Meta AI by using your voice, or use the button and touchpad on the side for basic controls like capturing images or videos and playing or pausing music. But unlike any other smart glasses to date, it also comes with an sEMG wristband in the box, Meta Neural Band.

In its current form, Meta Neural Band is set up to detect five gestures:

• Thumb to middle finger pinch: double tap to toggle the display on/off, single tap to go back to the system menu, or hold for quick shortcuts to the 3 menu tabs.
• Thumb to index finger pinch: how you "click".
• Thumb to side of index finger double tap: invoke Meta AI.
• Thumb swiping against the side of your index finger, like a virtual d-pad, which is how you scroll.
• Thumb to index finger pinch & twist: to adjust volume or camera zoom, as you would a physical volume knob.

While in the above clips I have my arms extended to illustrate the gestures, the beauty of sEMG is that you don't need to. Your hand can be at your side, resting on your leg, or even in your pocket. And it works even in complete darkness.

The gesture recognition is almost flawless, with close to 100% accuracy. The one exception is that rarely, if I'm walking it will fail to pick up a sideways directional swipe. But in general Meta Neural Band works incredibly well. The volume adjustment gesture, for example, where you pinch and twist an imaginary knob, feels like magic.

And whether I'm washing my hands, eating food, or driving a car, I have never found the display waking by accident, as the middle finger double-tap gesture only ever triggers when I'm intentionally making it. The only accidental activation I've encountered is that sometimes, if I'm tapping my phone with my thumb, Meta AI will trigger. It's hard to imagine solving this without the low-level access to the phone OS that only companies like Apple and Google have.

The Meta Neural Band gestures control the HUD interface, which looks much like that of a smart watch. It has 3 tabs, which you horizontally scroll between:

• The center home tab (the default) shows the date and time, your notifications, and a Meta AI button, with tiny shortcuts to active audio or navigation at the top.
• The right tab is the two-column applet library: WhatsApp, Instagram, Messenger, Messages, Calls, Camera, Music, Photos, Captions, Maps, Tutorials, and a game called Hypertrail. Four rows are shown at a time, and you navigate vertically to see the rest.
• The left tab features quick controls and settings like volume, brightness, Do Not Disturb, as well as shortcuts to Captions, Camera, and Music.

When I first used Meta Ray-Ban Display I found the interface to be "just too much", often requiring too many sequential gestures to do what you want. While I still broadly hold this view a month later, I have found myself becoming more used to quickly performing the correct sequence with experience, and I've discovered that if you pinch and hold your thumb to your middle finger, you get shortcuts to the three main menu tabs, which can speed things up.

I still think there's plenty of room for improvement in Meta Ray-Ban Display's interface though, from a simplicity perspective, and repeat my assertion that the menu should have two tabs, not three. The Meta AI finger gesture makes the Meta AI button on the center tab redundant, for example, and when you don't have any notifications, the tab feels like a waste of space.

A lot of the friction here will eventually be solved with the integration of eye tracking. Instead of needing to swipe around menus, you'll be able to just look at what you want and pinch, akin to the advantages of a touchscreen over arrow keys on a phone. But for now, it can sometimes feel like using MP3 players before the iPod, or smartphones before the iPhone. sEMG is obviously going to be a huge part of the future of computing. But I strongly suspect it will only be one half of the interaction answer, with eye tracking making the whole.

A major improvement since the Meta Connect demo though, is performance. While I still wouldn't describe Meta Ray-Ban Display as very snappy, the abject interface lag I encountered at Connect is gone in the shipping consumer model. The most noticeable delay is in waking the display, which often takes a second or two after the gesture.

Coming back to Meta Neural Band for a second, the only real problem with it is that it's something else you need to wear and charge (with a proprietary cable). I always wear a Fitbit on my left wrist, and now I wear the Meta Neural Band on my right too.

That's not to say that Meta Neural Band is a comfort burden. I find it no more or less comfortable than I did the Pixel Watch I used to own, and having tried a Whoop it feels similar to that too. And while it does leave a minor mark on my wrist, so do the straps of those other devices.

But it's another thing to remember to put on charge at night, another cable to remember to bring when traveling, and another USB-C port needed at my bedside.

Ideally, I should only have to wear and charge one wrist device. But today, Meta Neural Band is solely an sEMG input device, and nothing more.

The band already has an accelerometer inside, so in theory, a software update could let it track your daily step count. And if a future version could add a heart rate sensor for fitness, health, and sleep tracking, I wouldn't need my Fitbit at all anymore. But we're not there yet. And wearing a second wrist device just for input is a big ask for any product.

Features & Use Cases

There are six primary use cases of Meta Ray-Ban Display. It's a camera, a communications device, an on-foot GPS navigator, an assistive captions and translations tool, a personal audio player, and an AI assistant that can (if you want) see what you see.

So how well does it do each of these things?

Capturing Photos & Videos

When the regular Ray-Ban Meta glasses first launched, they were primarily pitched as camera glasses, like their predecessor the Ray-Ban Stories, and this remains one of the biggest use cases even as Meta now calls the product category "AI glasses". But without a display, you didn't get a preview of what you're capturing, nor could you check if the result was any good until it synced to the phone app. Sometimes you got lucky, and other times you failed to even frame your subjects, an issue not helped by the camera on almost all smart glasses being on a temple rather than centered.

Meta Ray-Ban Display also has 32GB of storage for media, and also syncs everything to your phone via Wi-Fi 6 when you open the Meta AI app. But the experience of capturing media is fundamentally better, because you get a live visual preview of exactly what's in frame. And with the Meta Neural Band, you can capture without raising your arm, as well as adjust the zoom on the fly by pinching your index finger to your thumb and twisting, the same gesture used to adjust the volume.

The camera quality isn't as good as your phone, given that the sensor has to fit into the temple of glasses. It appears to be the same camera as the regular Ray-Ban Meta glasses, and while it can produce great results in daytime, in low-light environments or with heavy zoom you'll get a relatively grainy output. You can see some sample shots and clips in my colleague Ian Hamilton's launch-week impressions piece.

Hands-On With Meta Ray-Ban Display & Meta Neural Band Across New York City

UploadVR’s Ian Hamilton bought the Meta Ray-Ban Display glasses on launch day and tested them across Manhattan.

UploadVRIan Hamilton

Regardless, the ability to capture precisely framed shots without needing to hold up a smartphone is Meta Ray-Ban Display at its best. It lets you both stay in the moment and capture memories at the same time, without the doubt that what you've shot doesn't capture what you wanted to include. And it works completely standalone, even if your phone is out of battery.

With a couple of swipes and taps you can also easily send your captured media to a friend, something that required extensive voice commands with the displayless glasses. And this brings us on to messaging.

Messaging

While Meta doesn't have an existing device ecosystem like Apple and Google, what it does have is the most popular messaging platform and two most popular social networks in the world, all three of which have an integration on Meta Ray-Ban Display.

You can opt to have incoming WhatsApp, Messenger, Instagram, and text messages pop up on the display, and so without needing to look down at a smartwatch or take your phone out of your pocket you can "screen" them to decide which are important enough to respond to immediately. Tap your thumb to your middle finger to dismiss a message, or to your index finger to open it.

The notification appears in the very bottom of the display area, which is already slightly below your sightline, so doesn't block your view of what's in front of you. And of course, unlike with a phone, no one around you can see it. There's also a setting to automatically detect when you're in a moving vehicle, so if you start driving a car you won't be interrupted.

If you want to respond to a message, there are 4 options in the interface: dictate, voice note, suggested emoji, or suggested reply.

I don't send voice notes, and I don't find the suggested emojis or replies useful, but I do love the dictation. It's powered by an on-device speech recognition model with relatively low latency and surprisingly great accuracy, in my experience. Even with my Northern Irish accent that some other systems (and people) can find difficult to understand, I'm able to dictate full responses without needing to type. And what's most impressive is that it works even when you speak quietly, thanks to the six-microphone array that includes a dedicated contact mic positioned just a few inches above your lips. That's yet another advantage of the glasses form factor.

Still, there are plenty of messages that I wouldn't want to dictate even softly in public, and situations when I want to use combinations of punctuation and letters that don't have a spoken form. Meta plans to release a software update in December that will let you enter text by finger-tracing letters on a physical surface, such as your leg, bringing some of its more advanced sEMG research out of the lab and into the product. It sounds straight out of science fiction, and we'll bring you impressions of it when the update rolls out. But it's not there today.

As well as receiving new messages as pop-up notifications, you can also access your 10 most recent threads on each messaging platform at any time by swiping to it on the apps list. In the apps, you can scroll through past messages and send new ones, including your captured photos and videos stored on the glasses.

The problem with accessing your past message threads, and with viewing photos and videos you've been sent, is that it's incredibly slow to load. Open WhatsApp on your phone and you'll typically see your messages update within a fraction of a second. On Meta Ray-Ban Display you'll see everything as it was the last time the applet was opened for upwards of 10 seconds, while media can take over a minute at times, or even just seemingly never load. And it's badly missing any kind of loading progress bar.

So, for example, while in theory you can definitely use Meta Ray-Ban Display to catch up on the dozens of Reels that one unemployed friend sends you all day on the go, assuming your eyes can put up with the monocular display for this long, in practice, as with waiting for fast-moving group chats to finally load, I often found it faster to take the phone out of my pocket and open the real app.

The cause of this slowness seems to be that Meta Ray-Ban Display is entirely reliant on your phone's Bluetooth for internet connectivity. This connection speed issue is something I ran into repeatedly on Meta Ray-Ban Display, and made me wish it had its own cellular connection. In fact, I'm increasingly convinced that cellular will be a hard requirement for successful HUD and AR glasses.

Audio & Video Calls

For over two years now, I've made and taken almost every phone call, both personal and professional, on smart glasses. I hate in-ear and over-ear audio devices for calls because it feels unnatural to not hear my own voice clearly, and the people on the other end love the output of the optimally-positioned microphone array.

With Meta Ray-Ban Display, the addition of the HUD and wristband lets you see how long a call has gone on for, and end it without having to raise your hand. You can also view and call recently called numbers in the Calls applet.

On the regular Ray-Ban Meta glasses you can also share your first-person view in a video call, and the big new calling feature of Meta Ray-Ban Display is that you can now see the other person too.

But again, this great-in-theory video calling feature is ruined by the fact that the internet connection is routed to Meta Ray-Ban Display via your phone's Bluetooth. Even with my phone and the person I'm calling on very strong internet connections, the view on both sides was pixelated and exhibited constant stuttering. Bluetooth just isn't meant for this.

On-Foot Navigation

One of the features I've most wanted HUD glasses (and eventually AR glasses) for is pedestrian navigation. There are few things I hate more in this world than arriving in a new city and having to constantly look down at my phone or watch while walking, almost bumping into people and poles. Worse, in cities with dense skyscrapers, the GPS accuracy degrades to dozens of meters, and I hopelessly watch the little blue dot on my phone bounce around the neighborhood.

In theory, Meta Ray-Ban Display solves at least the first problem, but with a massive caveat you absolutely must be aware of if you're thinking of buying it for this use case.

You can open the Maps applet anywhere, and you'll see a minimap (powered by OpenStreetMap and Overture) with nearby venues and landmarks. You can zoom all the way in to the street level, or out to the city level, using the same pinch-and-twist gesture used for volume control. You can also search for places using speech recognition, and scroll through the results by swiping your fingers.

The problem is that everywhere except inside the 28 cities Meta explicitly supports, you won't be able to initiate navigation. Instead, you just have an option to send it to your phone, where you can tap to open it in Google Maps, defeating the purpose. It's a rare product where core functionality is geofenced to a handful of cities.

I have been able to use Meta's navigation feature multiple times when visiting London, and found it genuinely very useful when in New York for the Samsung Galaxy XR launch event. I love the idea here, and where it works, the implementation isn't bad. Not having to look down to navigate is exactly what I wanted. But it works in so few places, relative to my life at least, that I just can't stop wishing it had Google Maps. And it's hard to imagine not switching to whatever HUD glasses have Google Maps first.

It's baffling that Meta decided to roll its own navigation system. It may be the right bet in the long term, but in the short term it compromises the product and leaves the goal wide open for Google to deliver a significantly better experience. Meta already has a wide-ranging partnership with Microsoft – why didn't it license Bing Maps for navigation? Or why not acquire a provider like TomTom?

It somewhat reminds me of the Apple Maps launch debacle, except in a parallel universe where it arrived alongside a first-generation iPhone that didn't have Google Maps.

As for the other problem with navigating in cities, the GPS issue, Meta Ray-Ban Display offers no solution. In theory, Meta could leverage the camera to calibrate the position and orientation, a technique often called VPS, but it doesn't today, and would likely require the company to build up a huge imagery dataset similar to Google Street View.

Meta AI & Its Dedicated Gesture

Meta has been marketing its smart glasses as "AI glasses" for some time now, riding the wave of hype that has for better and for worse almost entirely taken over the tech industry.

I didn't use Meta AI often on the regular Ray-Ban Meta glasses because I hate saying "Hey Meta", just as I hate saying "Hey Google", especially in public. With Meta Ray-Ban Display, you can still invoke the AI that way if you want, but there's also a dedicated gesture: just slightly curl your fingers inward and double-tap the side of your index finger with your thumb.

There's something very satisfying about this gesture. It feels more natural than the pinches used for everything else. And it has me using Meta AI far more often than I would if I had to keep saying "Hey Meta".

With the display, you also get visual aids in your responses. Ask about the weather in a place, for example, and you'll see a five-day forecast appear. Or for most queries, you'll see the response appear as text. I wish I could disable the voice response and just get the text, for some situations, and while I can do so by just reducing the volume to zero temporarily, this isn't a very elegant solution.

The real problem with Meta AI is that it's still Meta AI. It just isn't as advanced as OpenAI's GPT-5 or Google's Gemini 2.5, sometimes failing at queries where it needs to make a cognitive leap based on context, and fundamentally lacking the ability to think before it responds for complex requests.

Occasionally, I've run into situations where I wanted the convenience of asking advanced AI about something without taking out my phone, but ended up doing so after Meta AI just couldn't get the answer right. Ideally, I'd be able to say "Hey Meta, ask [ChatGPT/Gemini]". But that's not supported.

This is part of the reason Mark Zuckerberg is spending billions of dollars acquiring top AI talent for Meta Superintelligence Labs.

Audio Playback & Control

The primary way I use the regular Ray-Ban Meta glasses is for listening to podcasts and audiobooks. Smart glasses speakers don't do music justice, but they're great for spoken word content, and having your ears fully open to the real world is ideal.

What's different on Meta Ray-Ban Display is that you can far more easily and precisely adjust the volume. Rather than needing to raise your arm up to your head and awkwardly swipe your finger along the temple, you can just wake the display, pinch and hold your index finger to your thumb, and twist. It's a satisfying gesture that feels natural and precise.

The HUD also shows you a thumbnail for the content you're viewing, as well as how far you're into it and how long is left. It's a nice addition, and one less reason to take out my phone.

Live Captions & Translation

For accessibility, one of the biggest marketed features of Meta Ray-Ban Display is Live Captions.

The real-time speech transcription has decent accuracy, with the exception of niche proper nouns, and fairly low latency, always giving you at least the gist of what the person you're looking at is saying. And yes, I do just mean the person you're looking at. It's remarkable how well the system ignores any audio not in front of you, leveraging the microphone array to cancel it out. Look away from someone and the captions will stop. Look back and they'll continue. It really does work.

Just one swipe over from the virtual button to start Live Captions is the one to start Live Translation, and this feature blew me away. Having a friend speak Spanish and seeing an English translation of what they're saying feels like magic, and I could see it being immensely useful when traveling abroad. For the other person to understand you, by the way, you just hand them your phone with the Meta AI app open it shows them what you're saying, in their language. Brilliant.

Yes, you can do live translation with just a phone, or audio-only devices like AirPods and Pixel Buds.

Unfortunately, it only supports English, French, Spanish, and Italian. This is another example of where Google's services, namely Google Translate, would be immensely valuable.

The bigger problem with both Live Captions and Live Translation is that because the display is slightly below and to the right of your sightline, you can't look directly at someone when using the features. It's far better than having to look down at a phone, but ideally I'd want the text to appear centered and much higher, so that I could appear to be keeping eye contact with them while seemingly-magically understanding what they're saying. This would require different hardware, though.

The Big Missing Feature

While I'm glad that Meta Ray-Ban Display lets me decide whether it's worth taking my phone out of my pocket for inbound personal messages, what I want the most out of HUD glasses is the ability to screen my emails and Slack messages.

There is no applet store on Meta Ray-Ban Display, and Meta's upcoming SDK for phone apps to access its smart glasses won't support sending imagery to the HUD. For the foreseeable future, any new "apps" will have to come directly from Meta, and I call them "applets" because each really only does one thing and is essentially part of the OS.

(The company says it plans to add two new apps soon, a Teleprompter and a dedicated IG Reels experience.)

So a Slack applet won't be happening anytime soon. But there's a far easier way I could get what I want here.

Many smartwatches (yes, including third-party ones on iPhone) already let you view your phone notifications, so it's definitely technically possible. I asked Meta why it doesn't do this for Meta Ray-Ban Display, and the company told me that it came down to wanting to not overwhelm the user. I don't understand this answer though, since as with smartwatches, Meta could let you select exactly which apps you do and don't want notifications from, just as you can already for WhatsApp, Messenger, Instagram, and texts.

If I had this feature, letting me screen Slack notifications and emails, I would take my phone out of my pocket far less often. If any similar pair of glasses arrived with this feature, I'd switch over immediately.

The Case Folds Down & Has Huge Potential

Just like with AirPods, for smart glasses the included battery case is almost as important as the device itself. Meta Ray-Ban Display has an official "mixed use" battery life of 6 hours, which I've found to be fairly accurate, while the case provides 4 full charges for a total of 30 hours of use between needing to charge it with a cable.

The problem with the regular Ray-Ban and Oakley Meta glasses cases is that they're far too bulky to fit in almost any jacket pocket. And the Meta Ray-Ban Display case has an elegant solution for this, likely inspired by what Snap did in 2019 for the Spectacles 3.

When containing the glasses, the case is a triangular prism. But when not, it folds down into something with very similar dimensions to a typical smartphone, just slightly taller and less wide. This means that not only does it fit in most jacket pockets, but it even fits into my jeans pocket. So when I'm dressed relatively light and using Meta Ray-Ban Display as my sunglasses, for example, I can keep the case in my pocket when on-the-move and put the glasses back in the case when in the shade. The glasses and case together are the product, along with the wristband, and the folding design makes the whole system significantly more portable.

In fact, the glasses and case make such a great pair that I'm eager for the case to do more than just act as a battery and container.

When I need to take the glasses off to wash my face, take a shower, or just let them charge, docking them in the case should turn the duo into a smart speaker, letting me continue to listen to audio, take calls, and prompt Meta AI as I would with Alexa on an Amazon Echo. This could be achieved with no extra hardware, but ideally Meta would add a speaker to the case.

It also seems like the battery case could be the perfect way to bring cellular connectivity to a future version of Meta Ray-Ban Display without nuking the battery life of the glasses. The case would send and receive cellular signals, and relay the data to the glasses via Bluetooth for low-bandwidth tasks and Wi-Fi for high-bandwidth.

Interestingly, Meta has a partnership with Verizon to soon sell Meta Ray-Ban Display in stores. Could this evolve into a cellular partnership for the next generation?

Conclusions: Is It Worth $800?

Meta Ray-Ban Display is very much so a first generation product, an early attempt at a category that I'm convinced, after a month of testing, could eventually become a key part of the lives of at least hundreds of millions of people, if not billions.

In its current form, it's an amazing way to capture photos and videos without leaving the moment, and a great way to decide whether WhatsApp messages are worth your time without taking out your phone. In the cities where it's supported, the on-foot navigation is genuinely useful, and for the languages supported, the live translation feature could change what it means to travel. Like displayless smart glasses, it's also a great way to listen to spoken audio and take calls.

But the monocular display just doesn't feel right, too much of what Meta is trying to do is hampered by the device's lack of cellular connectivity, and the lack of established services like Gmail, Google Maps and Google Translate makes Meta Ray-Ban Display so much less useful than HUD glasses theoretically could be.

Further, while the Meta Neural Band works incredibly well, future versions need to replicate the functionality of smartwatches, instead of asking buyers to justify wearing yet another device on their wrist.

If you're an early adopter who loves having the latest technology and you don't mind looking odd in public, can live with the flaws I've outlined, and are fine with wearing a dedicated input device on your wrist, there's nothing else quite like Meta Ray-Ban Display, and the novelty could make up for the issues.

For everyone else, I recommend waiting for future generations of HUD glasses, ideally with binocular displays and either cellular connectivity or a seamless automatic phone Wi-Fi sharing system that I suspect only Apple and Google, the makers of your phone's OS, can pull off.

Just like with its Quest headsets, Meta is set to see fierce competition from the mobile platform incumbents in this space, and it'll be fascinating to see how the company responds and evolves its products through the rest of this decade and beyond.

Appreciate our reporting? Consider becoming an UploadVR Member or Patron.

The TL;DR

Reinforcement Learning from Human Feedback (RLHF) is the secret sauce that transforms powerful but useless AI models into the helpful, coherent assistants we use today like ChatGPT.

• LLMs are trained on the entire internet, which gives them a ton of knowledge…but not a clue how to use it

• RLHF refines them by showing them what we think a “good” answer looks like, teaching them how to be helpful assistants

• There are 3 steps to RLHF, starting with humans ranking different AI generated answers to common questions

• RLHF is critically important and a major focus of most AI labs today

LLMs before RLHF: absolute genius toddlers

If you were playing with AI models a few years ago, before the ChatGPT moment – or if you’ve ever accidentally used the “base model” of any of the popular models today – you may remember what life looks like without RLHF. You might ask a simple question and get back a wall of text that was technically correct but completely missed the point. Or it might hallucinate wildly, confidently invent facts, or worse, regurgitate some of the green toxic sludge it learned from the darker corners of the internet.

These early models were like a genius who had read every book in the world but had zero social skills. They had all the information, but they had no concept of what was helpful, polite, or relevant to a human – nor how to be concise, like at all.

To understand why this is true, you have to remember that LLMs are essentially random word generators that are trained to be really, really good at the word-guessing game. When you give them a prompt, they simply use their internal probability matrices to generate the most likely next word, most likely next word after that, and so on and so forth. This response is purely based on what appears most on the model’s internet-scale training set; the model has no idea how to be concise, helpful, or any of the other things we value in our assistant-like LLMs.

For example, GPT-2 once tried naming mushrooms and came up with things like:

So, how did we get from that useless genius to the polished, helpful assistant that can write you an email, a poem, or a Python script?

Before RLHF, we must understand RL

Before we add the “human feedback” part, let’s talk about Reinforcement Learning (RL) on its own. It’s a field of AI that’s all about teaching an “agent” (our AI model) to make good decisions to maximize a reward. And it has always been a bit of an odd step sibling of the more traditional methods of ML.

The easiest way to understand this is to think about training a dog. Imagine you’re teaching your new puppy (not a Doodle, please god) to fetch.

1. The Agent: The AI... I mean, the dog.

2. The Environment: Your backyard.

3. The Action: Mr. Dog can do lots of things—run, bark, dig up your prize-winning cilantro, or chase the ball.

4. The Reward: When Mr. Dog successfully brings the ball back, you give him a treat and say “Good boy!” That’s a positive reward. When he eats your herbs, you say “No!” That’s a negative reward (or lack of a treat).

Your dog isn’t a genius. He doesn’t understand physics or the emotional significance of your cilantro. He just learns, through trial and error, that one sequence of actions (chase ball -> pick up ball -> bring to human) leads to a precious treat. Over time, he’ll do that more and the other stuff less.

That’s Reinforcement Learning in a nutshell. You don’t give the model explicit instructions. You just give it a goal (get the reward) and let it figure out the best strategy. It’s how AI has learned to master complex games like Chess and Go, finding moves that no human had ever considered; and it is also the basis for RLHF.

Read more

Vous voulez utiliser Ollama sur votre iPhone ou Android pour lancer vos petits LLM en local ? Ce serait super cool non ? Bah j’ai une mauvaise nouvelle pour vous… votre smartphone n’a pas assez de mémoire vive pour faire ça…

Le problème est simple… les LLM bouffent un max de RAM. Par exemple, un LLaMA 7B, c’est dans les 12 GB de RAM. Et même quantifié en int4 pour gagner de la place, vous tombez à environ 3,5 GB. Et ça, c’est juste le modèle hein… Faut y ajouter le contexte, l’OS, les autres apps, et votre smartphone à 8 GB de RAM commence à suer de la raie.

Google a bien sûr sorti Gemini Nano pour Android, leur petit LLM optimisé pour mobile mais c’est compatible avec une poignée de smartphone car il faut un NPU dédié, assez de RAM, et une architecture très récente. Les autres, vous irez vous faire voir…

Du coup, une solution “pragmatique”, c’est de laisser votre Mac, votre PC, ou un petit serveur faire tourner Ollama chez vous, et d’utiliser votre smartphone comme simple client pour vous y connecter à distance. Vous gardez vos données locales, vous profitez de toute la puissance des modèles lourds, et votre iPhone ne chauffe pas comme un radiateur.

Et pour ça, il existe tout un tas d’apps mobiles qui font office de client Ollama. Des apps comme Enchanted sur iOS, My Ollama , Heat , et même les solutions cross-platform comme Ollamb codé en Flutter. Et aujourd’hui, je vous parle de Reins , une app développée par Ibrahim Cetin qui se démarque par des fonctionnalités que j’ai trouvées bien pensées.

Les fonctionnalités sont conçues pour des cas d’usage du monde réel. Vous pouvez ainsi définir un prompt system différent par conversation ce qui est pratique si vous avez un chat pour le code, un pour l’écriture, un pour la traduction et j’en passe… Vous éditez et régénérez les prompts à la volée et vous pouvez changer de modèle en cours de conversation sans tout perdre.

Reins supporte également l’envoi d’images, ce qui est utile si vous utilisez un modèle vision comme LLaVA. Vous pouvez aussi tweaker les paramètres avancés tels que la température, seed, taille du contexte, tokens max…etc. Et le streaming fonctionne en temps réel, comme ça pas besoin d’attendre une réponse complète avant de la voir.

Bien sûr, la question de la sécurité se pose. Il faut quand même exposer votre serveur Ollama sur Internet pour y accéder de l’extérieur donc pensez HTTPS obligatoire, tunnel ngrok temporaire si vous testez, VPN Tailscale ou Wireguard si vous voulez du permanent et sécurisé. Bref, les précautions classiques.

Vous pouvez télécharger Reins sur l’App Store, en APK ou récupérer les builds macOS/Linux sur GitHub .

Merci à Lorenper pour l’info.

L'intelligence artificielle va-t-elle remplacer l'humain pour gouverner ? De Tirana à Brighton, la politique expérimente l'IA et ravive la peur de l'« algocratie », un gouvernement par des algorithmes.

Le 30 octobre, l’opérateur anonyme qui se cache derrière le site Archive.is a posté un truc sur X. Pas un long message, hein, mais juste le scan d’une assignation en justice envoyée par le FBI daté du jour même, accompagné d’un seul mot : “canary”.

Si vous me lisez depuis longtemps, vous savez ce que ça veut dire. Un warrant canary , c’est une technique pour contourner les bâillons juridiques. Ainsi, quand une agence gouvernementale vous sert une assignation avec interdiction d’en parler, vous ne pouvez pas dire “hey les copains, j’ai reçu une assignation”. Par contre, vous pouvez publier régulièrement “je n’ai reçu aucune assignation”. Et le jour où vous arrêtez de publier cette phrase, tout le monde comprend que le canari est mort.

Sauf que là, le canari n’est pas mort. Il chante fort en publiant directement l’assignation elle-même.

Le document demande à Tucows, le registrar canadien qui gère les domaines Archive.is, Archive.ph et Archive.today, de balancer toutes les infos sur leur client : nom, adresse, numéros de téléphone, logs de paiement, tout. Le FBI a jusqu’au 29 novembre pour obtenir ces données et bien sûr, le document précise : “Vous êtes prié de ne pas divulguer l’existence de ce subpoena indéfiniment, car toute divulgation pourrait interférer avec une enquête en cours.”

Raté, lol.

Depuis 2012, le domaine archive.is est enregistré sous le nom de “Denis Petrov”, à Prague. Denis Petrov, si vous voulez, c’est un peu l’équivalent russe de Jean Dupont donc autant dire que c’est probablement pas son vrai nom. Et durant ces 13 dernières années, personne n’a réussi à identifier la vraie personne qui se cache derrière ce service utilisé par des millions de personnes chaque mois.

En 2025, maintenir un service web aussi gros en terme de visites, tout en restant complètement anonyme, c’est un exploit. Il faut des serveurs et il faut payer ces serveurs. Il faut gérer les DNS, les noms de domaine, les sauvegardes donc à chaque étape, il y a normalement une trace. Un paiement, une facture, une identité à vérifier. Et pourtant…

Archive.is, pour ceux qui ne l’utilisent pas, c’est un service d’archivage web à la demande. En gros, vous lui balancez une URL, et il vous crache un snapshot de la page. Un genre d’instantané figé dans le temps. C’est donc un peu différent de la Wayback Machine de l’Internet Archive qui crawle méthodiquement le web pour garder une trace longue durée. Archive.is, c’est du court terme, du rapide, du “j’ai besoin d’archiver cette page maintenant avant qu’elle disparaisse”.

Et les gens utilisent ce service pour plein de raisons. Par exemple, archiver un thread Twitter avant qu’il soit supprimé, sauvegarder un article avant qu’il soit modifié ou encore documenter une preuve. Mais là où il excelle c’est dans le contournement des paywalls.

Et c’est ce dernier point qui énerve l’industrie médiatique. En juillet de cette année, la News/Media Alliance a même réussi à faire fermer 12ft.io, un autre service de contournement de paywall. Le fondateur, Thomas Millar, avait créé son service pendant la pandémie après avoir constaté que, je cite, “8 des 10 premiers liens sur Google étaient derrière un paywall”. 12ft.io était hébergé chez un provider classique, avec un nom et une adresse… Il s’est pris une menace légale, et le service a du fermer.

Mais Archive.is, lui, résiste. Comment ? Hé bien parce qu’il n’y a personne à assigner. Pas de boîte. Pas de CEO. Y’a juste un fantôme qui paie ses factures et maintient les serveurs.

A titre perso, je comprends pourquoi les paywalls existent… Les médias doivent se financer et le journalisme de qualité coûte cher. Mais quelque part, je trouve ça quand même hyper triste humainement et professionnellement, d’enquêter, de prendre le temps d’écrire un super truc afin d’informer les gens, pour au final être lu uniquement par trois pelés et un tondu…

Mais bon, c’est pas vraiment ça qui intéresse le FBI.

Ce qui les dérange, c’est pas le paywall. C’est l’anonymat. Cette idée qu’on puisse opérer une infrastructure critique sur web sans identité vérifiable, ça ne passe plus. Les gouvernements veulent savoir qui fait quoi et cela même si c’est légal, même si c’est utile. L’anonymat est devenu une anomalie.

Et c’est là que le “canary” prend tout son sens car en publiant cette assignation, l’opérateur d’Archive.is fait deux choses. D’abord, il prévient tout le monde qu’il est dans le viseur mais il transforme aussi un document juridique confidentiel en acte de résistance publique. Le FBI voulait enquêter en silence et maintenant, tout Internet sait.

Le FBI chasse un archiviste, c’est à dire quelqu’un dont le métier est de faire des snapshots de ce qu’on trouve sur le web avant que ça ne disparaisse. Et là, il vient d’archiver sa propre disparition potentielle. Son message Twitter est déjà dans les archives d’Archive.is lui-même…

Tucows, de son côté, a confirmé qu’ils “respectent les procédures légales valides”, ce qui veut dire qu’ils vont probablement fournir les infos. Sauf que si l’opérateur d’Archive.is a réussi à rester anonyme pendant 13 ans, je doute qu’il ait laissé son vrai nom et son adresse perso dans les champs du registrar. Il a probablement utilisé des services d’anonymisation de domaine, des boîtes postales, des paiements en crypto. Bref, le FBI va peut-être obtenir des données, mais ça mènera probablement à un autre fantôme… On verra bien.

Quoiqu’il en soit, dans 10 ans, tous les services web devront avoir un humain identifiable et assignable en justice derrière. C’est le sens de la vie… et cette époque où on pouvait lancer un service en ligne sans donner son identité, c’est terminé. Archive.is est donc peut-être le dernier dinosaure de cette ère révolue où Internet était encore un peu sauvage, un peu anonyme, un peu libre…

Le canari chante. Mais pour combien de temps encore ? Ça personne ne sait…

Source .