J’sais pas si vous lisez des mangas de temps en temps mais si vous êtes à jour, vous avez peut-être envie de lire la suite, mais malheureusement, souvent c’est pas encore traduit en français. Alors vous 3 solutions… soit vous patientez, soit vous apprenez le japonais… Soit, soit…
Soit vous installez Koharu, un logiciel de traduction de mangas propulsé par IA. C’est hyper bien foutu puisque ça détecte automatiquement les bulles de dialogue, ça lit le texte japonais via OCR, ça efface proprement le texte original avec de l’inpainting, ça traduit le tout avec un modèle de langage aux petits oignons et ça replaque le texte traduit dans la bubulle.
Tout ça en quelques clics, évidemment, sinon ce serait pas drôle !
Le projet est développé par
mayocream
et c’est du 100% Rust avec une interface Tauri. Pour ceux qui ne connaissent pas,
Tauri
c’est un peu l’équivalent d’Electron mais en plus léger et plus performant. Le moteur d’inférence utilisé, c’est
Candle
de HuggingFace, ce qui permet de faire tourner des modèles IA localement sans avoir besoin d’envoyer vos data dans le cloud.
Côté modèles, Koharu embarque plusieurs outils spécialisés. Pour la vision par ordinateur, on a
comic-text-detector
pour repérer les bulles (avec
le petit modèle custom de mayocream
),
manga-ocr
pour la reconnaissance de caractères et
AnimeMangaInpainting
pour effacer proprement le texte original. Pour la traduction, c’est vntl-llama3-8b-v2 ou Sakura-GalTransl-7B-v3.7 qui s’y collent et c’est sans galère puisque ces modèles se téléchargent automatiquement au premier lancement.
Et Koharu supporte évidemment l’accélération GPU donc si vous avez une carte NVIDIA, vous pouvez profiter de CUDA et pour les fans d’Apple Silicon avec un M1 à M5, Metal est également supporté. Bref, ça dépote et le logiciel gère aussi la mise en page verticale pour les langues CJK (Chinois, Japonais, Coréen), ce qui est plutôt indispensable quand on traduit des mangas.
Les sources sont dispo sur
Github
et y’a des binaires pour Windows et macOS directement sur la page des releases. Pour les autres plateformes, faudra compiler vous-même avec Rust et Bun.
Voilà, si vous rêvez de traduire ce manga obscur qui dort au fond d’un forum japonais, Koharu va vous plaire. Et un grand merci à Lorenper pour l’info !
Generative AI and robotics are moving us ever closer to the day when we can ask for an object and have it created within a few minutes. In fact, MIT researchers have developed a speech-to-reality system, an AI-driven workflow that allows them to provide input to a robotic arm and "speak objects into existence," creating things like furniture in as little as five minutes.
Set to launch in 2026, XREAL Aura is the first pair of see-through AR glasses to run Android XR. Here are the impressions from my first hands-on.
During a recent meeting with Google, the company shared with me a range of updates relating to Android XR, including the first look at XREAL Aura. Unfortunately I wasn’t allowed to take any photos or videos during the demo.
XREAL Aura will be the first pair of see-through glasses running the full-blown immersive version of Android XR. Considering the immersion they offer, the glasses are impressively compact. Much of this is thanks to offloading weight, battery, and compute onto a tethered puck which can slip into your pocket. Interestingly, the puck looks like the size and shape of a typical smartphone, but instead, the entire screen area is a giant trackpad which can be used for mouse-like input (in addition to hand-tracking).
Image courtesy Google
Compared to prior models of XREAL glasses with bird-bath style optics, the optics in the XREAL Aura are a bit more compact, allowing you to bring your eyes closer to the lens. XREAL is finally getting close to AR glasses which actually look like glasses without the lenses being so far from your eyes that it looks goofy. They’re still look a little funky because of the distance, but we’re getting there.
Putting on the glasses, I was presented with the same Android XR experience I’m used to from Galaxy XR, except this time the background I was seeing was the real world (albeit, a fairly dim version of the real world due to the light loss through the lenses). Unfortunately, Aura doesn’t include eye-tracking, which means I couldn’t use my preferred look+pinch method of input (which is supported on Galaxy XR); I had to use the default laser-pointer input style which feels more cumbersome.
The quoted 70-degree field-of-view (I assume diagonal, but they didn’t specify) felt usable, but seems like the bare minimum field-of-view needed to get real value out of an immersive operating system like Android XR. If Google hasn’t already specified that any immersive Android XR device needs at least this wide of an FoV, they absolutely should.
I was impressed with the sharpness and brightness of the Aura’s display. It looked pretty good for virtual screen usage (ie: viewing websites, videos, photos, etc). But I also saw obvious pupil swim (which looks like warping as you move your head around). This will be more noticeable in fully immersive experiences or multi-tasking where there’s lots of head movement. For some people, pupil swim is just a visual annoyance, but for others it can be dizzying over time. It’s unclear if this can be improved before launch, especially without any on-board eye-tracking hardware.
One particularly cool feature of Aura is the electronically-controlled dimming lenses. A button on the stem of the glasses allows you to dim the real world from 0% to nearly 100%, blocking out almost all light.
This isn’t the first pair of AR glasses to include electronic lens dimming, but the way it’s integrated into Android XR is a clever value-add. If you launch a fully immersive application (like a VR game), the software can automatically switch the dimming to 100% so the virtual content doesn’t conflict with your real-world background. Or, in applications that would like to dim the user’s background (like a media app), it can set the dimming to 50%. With a passthrough headset like Galaxy XR, this dimming would normally be done digitally, but with the Aura glasses it’s done physically. And Google says developers don’t need to worry about the distinction; if their app asks Android XR to dim the background, it’ll automatically do so through whatever means are available to the device.
Of course, Aura is a ‘glasses-style’ device, so even when dimming is set to 100%, you’ll still see lots of the real-world in your peripheral vision. But still, having the feature makes fully immersive applications usable in a way they wouldn’t on a see-through AR device.
Considering its capabilities, Aura looks strikingly close to a normal enough pair of sunglasses that you might not get a double take by people passing by, though anyone talking to you face-to-face would surely know there is something strange going on behind the lenses.
Even with dimming set to 0%, the real world is still made fairly dim; like wearing sunglasses inside. Whether by design (to account for not enough display brightness) or happenstance (as a result of light loss from the optics), it still feels like you’re wearing sunglasses inside, which limits some of the indoor use-cases you might want to do with a pair of AR glasses. For instance, I’d like to have AR glasses which work as a cooking companion in the kitchen so I can reference recipes while preparing food. But like a regular pair of sunglasses, Aura dims the world too much that I wouldn’t want to use them in the kitchen.
I look forward to trying the finished version. In my brief hands-on, I came away feeling like Aura is the first clear look at the eventual convergence of AR and VR headsets. It feels like a full-fledged Android XR headset but in a much more compact package that will be way more portable and less conspicuous. I could actually see myself using Aura on a plane or in a coffee shop without feeling like everyone would be staring at me.
Indeed, Google is thinking the same. Alongside my look at Aura, the company also announced that it’s rolling out a first-party PC Connect application for Android XR to make it easy to stream your Windows desktop to the glasses for productivity, media, or gaming.
There’s still some key things we don’t know about Aura. The company hasn’t revealed a full set of specifications yet, and we don’t know if there will be any controller support (which would mean incompatibility with many immersive VR games). We also don’t have a price or specific release date, but XREAL has confirmed that Aura will launch in 2026.
Rare il y a encore quelques années, le recours à des solutions de validation automatique d'identité est aujourd'hui incontournable dans les processus d'enrôlement « digital » des acteurs de la finance. Mais elles sont généralement issues de partenaires externes, à qui sont donc transmises des données sensibles. Sauf avec OCR Studio.
Vous connaissez maintenant le principe, que vous avez probablement déjà rencontré : au moment de, par exemple, ouvrir un compte bancaire depuis votre téléphone, vous devez d'abord présenter une pièce d'identité officielle. Une fois son authenticité confirmée, les informations d'état-civil en sont extraites, tout comme la photographie, qui est, elle, utilisée afin d'effectuer une vérification de la personne réalisant l'opération, grâce à une comparaison avec une capture vidéo pilotée sur son appareil.
Jusqu'à présent, ces fonctions sont, sauf exception, fournies par quelques entreprises spécialisées qui exécutent l'ensemble des traitements requis sur leurs propres infrastructures, souvent infonuagiques, où les images des pièces d'identité et les flux vidéo doivent donc préalablement être transférés. Entre autres défauts, un tel mécanisme a un impact sur les performances et peut soulever des inquiétudes sur la confidentialité des données (depuis les usages abusifs jusqu'aux risques de fuite).
La jeune pousse émiratie a une autre idée : avec sa nouvelle boîte à outils, toutes les étapes du processus sont implémentées directement sur le matériel du client (y compris, apparemment, via un navigateur web sur un micro-ordinateur). Son module d'intelligence artificielle spécialisée, fonctionnant sur un processeur classique, assure ainsi la reconnaissance des documents d'identité (plus de 200 modèles sont reconnus, couvrant le monde entier) et l'analyse de ressemblance du visage de l'utilisateur.
OCR Studio vante naturellement la rapidité de traitement de son dispositif – obtenue à la fois par la qualité de ses algorithmes et l'absence d'échanges potentiellement volumineux avec un serveur distant – et son engagement pour la protection de la vie privée – aucune information personnelle n'étant jamais divulguée. Une version utilisable en agence est également disponible, pour laquelle la promesse est équivalente, de conserver les données capturées dans l'environnement de l'établissement.
Ces avantages résonneront bien sûr auprès des clients finaux, en particulier au niveau de leur expérience, la réduction des temps d'attente constituant un facteur majeur de baisse des taux d'abandon. Mais la garantie de confidentialité devrait également séduire les départements de conformité des institutions financières, pour lesquels elle peut soulager des craintes vis-à-vis de prestations externalisées… sans parler des responsables obsédés par l'impératif de garder la mainmise sur les outils déployés.
Yesterday was the first day of United XR Europe. The Expo was not open (it opens on the second day), but being the pirate I am, I was still able to try a couple of experiences…
VIVE Eagle Smartglasses
Trailer of the Vive Eagle glasses
HTC VIVE does not have a booth at United XR Europe, but a Vive employee is going around with the glasses on, letting people try them. A guy strongly suggested to me to try this device, so I went to the VIVE afterparty, ate a lot of free food (which is the most important thing), and then, as soon as I saw a guy from HTC, I jumped on him, asking him if he was “the glasses guy”. He was not, but he took me to the right glasses guy, who was kind enough to let me try the glasses. The test was very brief and happened during a party, so what I’m going to write doesn’t count as a review and not even as a first impression… it is like “very early impressions”.
Anyway, the glasses I’ve tried came in two colors: a pair was red with transparent lenses, and the other was black with lenses that could get obscured. These are some pictures I’ve taken of the latter.
And this is a picture of me, my buddy Christian Steiner, and HTC’s Thomas Dexmier, from which you can see how they fit the faces of people.
Looking cool with the Vive Eagle smartglasses
The glasses have a nice design, but it is not an elaborate one: so I look pretty good in the pictures, but I don’t seem like someone fashionable. I still prefer Ray-Ban’s design. What Vive did well was try to propose glasses with bold colors: Vive Eagles could be red or blue. I guess it is the only smartglasses brand doing that now, so we need to owe them they are trying something different. Bold colors only suit bold people, though: I don’t think everyone wants to go on the street with straight red glasses on their face, so it’s good they also have more traditional options like the black ones.
Still looking good with the black glasses
The glasses really look like regular glasses, and the only thing that makes them a bit different is that the temples of the frames are bulkier than usual. Wearing them, they felt lightweight and comfortable, even if they were a bit too heavy on my nose because this was the model currently sold in Taiwan and the surrounding areas. Chinese units are never comfortable for us Westerners, so for sure, HTC will slightly modify the fit of the device when selling it here in Europe. I would need more time to evaluate the exact comfort, but let’s say that for now it seemed pretty good… even if, as a non-glasses wearer, my face was constantly perceiving there was something new on my face.
Looking like a creative guy with the red model
Regarding the functionalities, these glasses have a camera and some integrated speakers. The HTC representative used his phone (with a companion app) to play some music on the device. Notwithstanding the messy and noisy space of the party, I could still clearly hear the song. This means the speakers are good.
The glasses companion app running on the (big) phone
I came out with a good impression from this quick test on the glasses. But the problem is that their price is NT$15,600, which is more than $500. These glasses have more or less the same features as Ray-Ban Meta and cost almost twice as much. Unless HTC offers a more competitive price in the West, it’s hard for it to compete…
Zoo Of The Future
The Zoo Of The Future is a location-based VR experience that is available within 5 minutes on foot from the UnitedXR Europe venue. The concept is pretty interesting: instead of taking the animals from their natural habitat and keeping them “prisoners” in a zoo to let people see them, people could visit them in their natural environment in virtual reality.
When you arrive at the Zoo Of The Future, after a brief registration, you are given headphones speaking the language you choose. Then you enter a closed room where there is projection mapping on the four walls. A video starts and basically introduces you to the experience, telling you why it has been created, while showing you different animals in their habitat. This video experience is already nice because it feels a tiny bit like a CAVE experience, where you are immersed in a place not via a VR headset, but via projections on the walls. This part ends with the narrating voice telling you that you can experience the animals via virtual reality. Remember that this is an experience for ordinary people, so it is important to explain things and let them slowly get used to virtuality (first with projections and then with a headset).
As you can see, the video projections are pretty well made
After this introductory room, people can enter the real “zoo”. The zoo is a physical space that is divided into zones, where every zone represents a different habitat, like Antarctica, the forest, or the savannah. Zones show different physical settings: for instance, Antarctica shows images of snow, and there is, for instance, a fake snow block there in the room. I’ve been told that every zone also has a slightly different smell, so that when you walk across the zone, your brain truly perceives you’re going into different environments.
Welcome to the jungle
You can interact with these different areas both in AR and VR. The AR interaction is pretty simple and mostly studied for kids: you’re given a tablet, and you can frame the picture of an animal on the walls to make it appear in AR. For instance, you look at a panda, and the animated 3D model of a panda eating bamboo appears in front of you. You can then take a picture of it. It is pretty cute and I’ve been told many people are also taking selfies with the AR animals.
This is the section of the location dedicated to Antarctica
Regarding VR, there are different stations with a display and a headset (currently a Meta Quest). When you put on the headset, you see yourself in the virtual environment where the animals are. For instance, if you were in the physical savannah, you could see a family of giraffes walking towards the water. You can not interact with animals: you are just an observer of them in their natural habitat. The only thing you can do is teleport around to discover the various animals inhabiting that place. To move around and teleport, you just point at locations with your index finger (the experience uses hand tracking to avoid messing with controllers handled by the visitors). The teleporting with the straight out index finger is a bit uncomfortable in the long run, but it has been chosen because it is easy to understand and so ideal for the average user. Also, the fact that as soon as you put the headset on, you are immediately in the experience (no menus, no intros, etcc…) has been done for the sake of simplicity.
One of the stations with the VR headset and the screen doing the mirroring
The graphical quality of the experience is definitely good. It has been made with Unreal Engine, and the animals are modeled pretty well. When I saw the giraffes in front of me, they looked pretty realistic. The used headsets are Quest 3 connected via USB to a computer: using PCVR, you can have pretty good graphics, so the experience feels immersive.
Is this like going to the zoo? Well, not for now, but I see the potential. The graphics are immersive but are not perfectly photorealistic; these are not some Gaussian splats recordings. The experience could compensate with something else, like letting you interact with the animals, or creating some missions you have to accomplish in the places you go, or fostering multiplayer interactions. The good thing about virtual reality is that you can do more than you can do in reality, and the Zoo Of The Future should exploit this more.
Giraffes were pretty nice to experience
The concept is for sure interesting, and with the continuous advancements of virtual reality, it could in the future really feel like being with the animals in their habitats, making the zoos mostly obsolete. And the Zoo Of The Future is a good present representation of that: I think the team working on it did a great job both with the physical space, and the easy-to-use AR and VR experiences. If you are around UnitedXR Europe, I suggest you visit it: during the event, there is a heavy discount, and it is just 5€ per person. You can find more information here: https://zooofthefuture.be/
A Panel on the main stage
Today I also had my panel together with Gabriele Romagnoli, Casandra Vuong, and David Heaney, moderated by Peter Siekerman.
A random selfie I shot before we got on stage
It was nice having a chat with these smart people on stage. And for me, it also also an important achievement because it was the first time on the main stage of an event of the AWE family. I’ve always spoken in secondary rooms, so it felt a bit special to me, and that’s why I took a selfie from the stage after we finished our talk! Probably it is not going to happen again… but who cares, I managed to have at least my moment of glory!
MAAAAAAIN STAAAAAGE!
Ah, and Ori Inbar made me scream during his keynote, but that’s another story for another day…
[Voria Labs] has created a whole bunch of artworks referred to as Lumanoi Interactive Light Sculptures. A new video explains the hardware behind these beautiful glowing pieces, as well as the magic that makes their interactivity work.
The basic architecture of the Lumanoi pieces starts with a custom main control board, based around the ESP-32-S3-WROOM-2. It’s got two I2C buses onboard, as well as an extension port with some GPIO breakouts. The controller also has lots of protection features and can shut down the whole sculpture if needed. The main control board works in turn with a series of daisy-chained “cell” boards attached via a 20-pin ribbon cable. The cable carries 24-volt power, a bunch of grounds, and LED and UART data that can be passed from cell to cell. The cells are responsible for spitting out data to addressable LEDs that light the sculpture, and also have their own microcontrollers and photodiodes, allowing them to do all kinds of neat tricks.
As for interactivity, simple sensors provide ways for the viewer to interact with the glowing artwork. Ambient light sensors connected via I2C can pick up the brightness of the room as well as respond to passing shadows, while touch controls give a more direct interface to those interacting with the art.
[Voria Labs] has provided a great primer on building hardcore LED sculptures in a smart, robust manner. We love a good art piece here, from the mechanical to the purely illuminatory. Video after the break.
At CES 2026, Lenovo going all-in on AI (like every other company), but it’s hiding all the fun goodies in the back, and I’m excited about what might be coming.
Issu de l’ETH Zurich, l’une des principales universités mondiales en ingénierie et en robotique, Mimic Robotics s’attaque à l’un des derniers bastions de l’automatisation industrielle, celle des tâches manuelles complexes qui exigent une dextérité humaine réelle. La startup suisse part du constat que les robots traditionnels excellent dans les gestes répétitifs, mais échouent dès qu’il …
Researchers at UC Santa Barbara have invented a display technology for on-screen graphics that are both visible and haptic, meaning that they can be felt via touch.
What was once a healthy lead over its competition thanks to its blockbuster AI chatbot ChatGPT has turned into a razor-thin edge, motivating OpenAI CEO Sam Altman to declare a “code red.”
The financial stakes are almost comical in their magnitude: The company is lighting billions of dollars on fire, with no end in sight; it’s committed to spending well over $1 trillion over the next several years while simultaneously losing a staggering sum each quarter.
And revenues are lagging far behind, with the vast majority of ChatGPT users balking at the idea of paying for a subscription.
Meanwhile, Google has made major strides, quickly catching up with OpenAI’s claimed 800 million or so weekly active ChatGPT users as of September. Worse yet, Google is far better positioned to turn generative AI into a viable business — all while minting a comfortable $30 billion in profit each quarter, as the Washington Post points out.
The question on many investors’ minds: if the AI bubble were to collapse, would OpenAI even survive?
“We’re going to see a situation where ChatGPT was the early winner,” Porter and Co. equity analyst Ross Hendricks told WaPo. “They’re going to end up just like MySpace did with the inability to truly monetize and break away from the pack.”
In a Thursday note, Deutsche Bank analyst Jim Reid estimated staggering losses for OpenAI amounting to $140 billion between 2024 and 2029.
“OpenAI may continue to attract significant funding and could ultimately develop products that generate substantial profits and revolutionize the world,” he wrote, as quoted by WaPo. “But at present, no start-up in history has operated with expected losses on anything approaching this scale.”
“We are firmly in uncharted territory,” Reid added.
To even just cover the interest on all the money OpenAI is borrowing, the company’s revenues will have to grow immensely. But according to recent Sensor Tower data reviewed by WaPo, ChatGPT’s monthly active users grew by a mere five percent between July and November. Google’s Gemini AI app soared by a far healthier 30 percent over the same period.
Recent data also suggests ChatGPT user growth is stalling out in Europe, highlighting a slowdown that couldn’t come at a worse time for OpenAI.
Google’s latest Gemini 3, in particular, impressed when it was announced released last month, with benchmarks exceeding OpenAI’s most powerful AI models. Its Nano Banana Pro AI image model has also pushed the envelope, while OpenAI’s Sora video-generating app has received comparatively little media attention after a storm of controversy surrounding its rollout.
It’s not just Google, either. OpenAI is also facing steep competition from open-source AI models in China, like the startup DeepSeek, whose extremely energy-efficient R1 model threw Silicon Valley into chaos earlier this year.
In short, by many indications, OpenAI appears to be in deep water, and analysts are growing wary of the company burning through astronomical amounts of money at an unprecedented pace.
Even the so-called “Godfather of AI” and former Google AI lead, Geoffrey Hinton, isn’t optimistic about OpenAI’s future.
“I think it’s actually more surprising than it’s taken this long for Google to overtake OpenAI,” he told Business Insider this week.
“I think that right now they’re beginning to overtake it,” he added. “Google has a lot of very good researchers and obviously a lot of data and a lot of data centers.”
Pour la première fois en France, Tesla a ouvert l'accès à sa technologie de Conduite Entièrement Autonome surveillée (FSD) au grand public et à la presse. En décembre, une dizaine de véhicules équipés du dernier matériel (Hardware 4) et du logiciel FSD v14 ont sillonné les rues françaises. Loin des autoroutes californiennes et des villes américaines au quadrillage parfait, comment l'IA s'en sort-elle face aux ronds-points, aux travaux et à l'imprévisibilité des routes européennes ? C'est ce qu'on va voir.
Depuis son orbite à 540 kilomètres d'altitude, Hubble scrute l'univers depuis plus de trente ans. Mais voilà qu'un nouveau problème surgit : les satellites de télécommunications traversent de plus en plus souvent son champ de vision. Entre 2019 et aujourd'hui, leur nombre est passé de 2 000 à 15 000 en orbite basse. Ce n'est qu'un début.Quand les tél...
Un agent du help desk décroche un appel, la voix se présente comme celle de Sara, paraît pressée, évoque un incident récent, cite des projets internes tout en adoptant les tournures familières que Sara emploie d’ordinaire. Elle sollicite une simple réinitialisation de mot de passe, rien ne semble inhabituel, pourtant, ce n’est pas Sara au …
Microsoft has lowered sales growth targets for its AI agent products after many salespeople missed their quotas in the fiscal year ending in June, according to a report Wednesday from The Information. The adjustment is reportedly unusual for Microsoft, and it comes after the company missed a number of ambitious sales goals for its AI offerings.
AI agents are specialized implementations of AI language models designed to perform multistep tasks autonomously rather than simply responding to single prompts. So-called “agentic” features have been central to Microsoft’s 2025 sales pitch: At its Build conference in May, the company declared that it has entered “the era of AI agents.”
The company has promised customers that agents could automate complex tasks, such as generating dashboards from sales data or writing customer reports. At its Ignite conference in November, Microsoft announced new features like Word, Excel, and PowerPoint agents in Microsoft 365 Copilot, along with tools for building and deploying agents through Azure AI Foundry and Copilot Studio. But as the year draws to a close, that promise has proven harder to deliver than the company expected.
A new study shows AI and social platforms are driving record traffic to online stores—but 80% of shoppers still abandon their carts Online retail is experiencing rapid traffic growth, yet most shopping journeys end without a purchase. To understand why, Semrush and Statista analyzed millions of shopping sessions across major e-commerce platforms. Their report, “Rewriting […]
For many months now, AI companies have made a huge deal out of “AI agents,” meaning autonomous software systems that can make decisions and take actions on behalf of humans with minimal intervention.
But when that ambitious vision will turn into a reality remains anybody’s guess. The current crop of agentic AI models is still getting easily tripped up, often requiring humans to jump in, effectively undercutting their purpose.
The numbers remain dismal. Researchers at Carnegie Mellon University found earlier this year that even the best-performing AI agent, which was Google’s Gemini 2.5 Pro at the time, failed to complete real-world office tasks 70 percent of the time.
And this summer, OpenAI released its ChatGPT agent, promising that it can “do work for you using its own computer, handling complex tasks from start to finish.” But in reality, users found the experience lackluster, calling it “not very useful,” “shaky,” and “slow.”
As such, it shouldn’t come as much of a surprise that Microsoft is struggling to sell its enterprise clients on its own take on agentic AI. As The Information reports, the company’s Azure salespeople are seriously struggling to meet some extremely ambitious sales growth targets, cutting quotas by up to 50 percent earlier this year.
Microsoft’s stock sank over 2.5 percent on Wednesday, showing that investors are unimpressed by the tepid sales results.
Meanwhile, the tech giant went into damage control mode, with a spokesperson telling Bloomberg in a statement that “The Information’s story inaccurately combines the concepts of growth and sales quotas” and that “aggregate sales quotas for AI products have not been lowered.”
Regardless, the dustup suggests that enterprise customers are far from convinced that large AI agents are ready to autonomously complete complex multistep tasks. It’s yet another indication that companies are struggling to convert the enormous hype surrounding generative AI into actual revenue, a concerning trend considering the billions of dollars AI companies are burning through right now with no end — or return on investment — in sight.
It’s not like Microsoft’s clients are being unreasonable. Generative AI continues to struggle with the absolute basics, and hallucinations remain a major pain point. Multiply the potential for an AI to make up facts as it attempts to complete a more nuanced, multistep project, and the chances of it tripping up rise even further.
In short, the future that’s being sold to these customers simply hasn’t materialized. And that could hamper AI companies’ sky-high expectations when it comes to monetizing the tech.
Then there’s competition ratcheting up the pressure for Microsoft. In June, Bloomberg reported that workers preferred to use OpenAI, which was cutting into its ability to sell its Copilot.
Fortunately for Microsoft, most of its current revenues come from renting out cloud computing infrastructure to AI companies, not selling AI products to enterprise customers, as The Information notes.
Nonetheless, the cracks are starting to show, indicating sales could continue to lag behind goals as customers realize they might be being sold a vision of a distant future.
Albertsons Companies has launched an AI shopping assistant, a web browser experience designed to make grocery shopping faster, smarter and more personalised.
Building on the company’s Ask AI tool introduced earlier this year, the agentic shopping assistant is powered by various collaborative agents and is now available on all Albertsons Cos. banner websites as Albertsons AI, Safeway AI, Vons AI, Jewel-Osco AI etc.
“Our goal is to make our customers’ lives easier, and by implementing AI powered features across the customer journey from discovery to purchase, we are delivering an experience that’s faster, easier and more enjoyable,” says Jill Pavlovich, SVP, Digital Customer Experience at Albertsons Cos.
“We are laser focused on using AI as part of our strategy to meet customers when and how they choose to shop, ultimately driving customer growth and engagement through digital connection. The Albertsons AI shopping assistant is an exciting step in this journey, with much more innovation to come.”
Powered by Open AI models, the Albertsons AI shopping assistant aims to reduce grocery shopping time from an average of 46 minutes to as little as four minutes by helping customers plan meals, restock essentials, discover new products etc. The assistant can be accessed on banner web browsers in the Meals Hub
“With the Albertsons AI shopping assistant, we can understand our customers’ needs on a deeper level, allowing us to serve them in the best way possible,” says Pavlovich.
“Our current meal-planning capabilities are exceptional, and now, through dynamic two way conversations with our shoppers, they can give us further context such as the customer is cooking a meal for six people but wants leftovers and one person is a vegetarian.”
“We can provide tailored recommendations in the moment with all of these considerations and enhance future interactions with personalised product and recipe recommendations featuring our latest deals and coupons.”
Expanding into the Albertsons Cos. banner mobile apps in early 2026, the shopping assistant will launch additional agentic commerce capabilities including budget optimisation, in-store aisle location to find a specific product, and voice integration. Additionally, ita multi‑agent architecture sets the stage for compatibility with off‑platform agents, opening doors for future integrations with experiences like apps in the chat feature.
As we witness a digital transformation revolution across all channels, AI tools are reshaping the omnichannel game, from personalising customer experiences to optimising inventory, uncovering insights into consumer behaviour, and enhancing the human element of retailers' businesses.
With 2025 set to be the year when AI and especially gen AI shake off the ‘heavily hyped’ tag and become embedded in retail business processes, our newly launched awards celebrate global technology innovation in a fast moving omnichannel world and the resulting benefits for retailers, shoppers and employees.
Our 2025 winners will be those companies who not only recognise the potential of AI, but also make it usable in everyday work - resulting in more efficiency and innovation in all areas.
Winners will be announced at an evening event at The Barbican in Central London on Thursday, 29th January. This will kick off with a drinks reception in the stunning Conservatory, followed by a three course meal, and awards ceremony in the Garden Room.
Latitude and longitude together give you your position on the globe. They are the two types of lines that wrap the planet on globes and our world maps. One set is the horizontal lines, and the other vertical, but for the longest time I struggled to remember which was which. No more.
How to Remember Latitude vs Longitude
The mnemonic I've found that has never failed for me to remember which is latitude and which is longitude is:
Latitude is flat. (or just Lat is flat)
Longitude converges.
So lines of latitude are the lines that lie flat on a globe or map. Lines of longitude are the tall ones that converge at the poles (latitude lines don’t converge).
Two other ways that might work for you:
Imagine the map grid as a ladder. Latitude lines are the flat rungs you could step on, and longitude lines are the long vertical lines connecting the rungs.
Longitude lines are always long, as they all pass through the poles. Whereas latitude lines, as they encircle at different positions north-south around the planet, are long by the equator but short around the poles.
Which Way Latitude and Longitude Run (and What They Mean)
When you give a latitude for a city or location, you're saying which of the horizontal lines of latitude the location is. Somewhat counterintuitively then, this tells you how far up or down the planet you are between the poles—your north–south position.
Latitude behaves like the y position on a graph. Latitude values increase as you go north from the equator and decrease as you go south.
When you give a longitude, this corresponds to which of the tall vertical longitude lines you mean. Longitude tells you where you are east–west around the world.
Longitude behaves like the x position on a graph. Longitude values increase as you go east from the prime meridian at Greenwich in London, and decrease (-ve) as you go west.
So:
Lines of latitude are horizontal, but latitude gives you your vertical position (y position)
Lines of longitude are vertical, but longitude gives you your horizontal position (x position)
Which comes first: latitude or longitude?
Most consumer tools like Google Maps write coordinates as lat, lon. An easy way to remember which comes first is alphabetical: latitude, then longitude—a then o.
So a GPS coordinate typically corresponds to your latitude (north–south) first and your longitude (east–west) second. Watch out for this if you, like me, are used to thinking of coordinates as x, y—horizontal position first, then vertical position.
Although many consumer tools use lat, long, some technical formats use long, lat, so it’s always worth checking. It can be a real pain if you travel to the wrong location.
A Lat/Lon Example
To take a real coordinate example. The coordinates for The British Library in London, written in the way Google Maps expects (latitude, longitude) are:
51.53005,-0.12765
That is 51.53005 degrees north of the equator, and just a tiny bit west of the Greenwich Prime Meridian.
You can type this into a search in Google Maps or Apple Maps to go straight there giving you:
Brace yourselves, for your cozy little conversations with OpenAI’s ChatGPT are about to be stuffed with ads.
At least, that’s the findings of a software engineer who dug through the code of an experimental build of the chatbot’s app.
The engineer, Tibor Blaho, flagged roughly a dozen lines of code in the latest beta release of the ChatGPT app for Android, 1.2025.329, labelled “feature ads,” with strings referencing commercial-sounding stuff like “search ad,” “search ads carousel,” and “bazaar content.”
It’s yet another sign that OpenAI is preparing to cash in on its hundreds of millions of users by showing them ads, amid growing pressure on the company to demonstrate it can make a profit while its spending — and AI capital expenditures at large — continue to balloon past what subscriptions can support.
OpenAI leadership has played coy when asked about ads. Last December, chief financial officer Sarah Friar told the Financial Timesthat the company, now valued at around half a trillion dollars, was exploring an ad model, but then backtracked by stating there were “no active plans to pursue advertising.”
CEO Sam Altman has sent similarly mixed messages. He once called the idea of integrating ads into ChatGPT “unsettling” and described them as a “last resort.” But this year, Altman said he wasn’t “totally against” advertising, that it was something he expected to “try at some point,” and praised Instagram’s ad model (which, it’s worth noting, has attracted an immense amount of criticism from non-CEOs.)
The warning signs have been there, however, regardless of the execs’ messaging.
Amid Friar’s waffling on the issue, for instance, the Financial Times reported that OpenAI was poaching top ad talent from its rivals like Google and Meta, while posting ad-related job listings on LinkedIn.
And last month, The Information reported that OpenAI was considering showing individually-tailored ads based on ChatGPT’s memory of user interactions. The reporting noted that the AI startup had imported hundreds of ex-employees from Meta, an advertising juggernaut; in 2023, 98 percent of Meta’s over $130 billion revenue came from ads.
If ads do come to OpenAI’s products, the change will be a controversial one. ChatGPT has been completely ad free since its launch three years ago, becoming part of its appeal as an alternative to ad-stuffed search engines like Google. Moreover, the intrusion of ads could dispel some of the human-like familiarity that users have fostered with the chatbot.
It will also raise a host of ethical concerns. Users tend to share much more intimate details with a chatbot than they do a search bar, and the AI’s ability to act like a friend and a confidante — traits that are themselves controversial amid more and more accounts of so-called “AI psychosis” — could be repurposed for sleazy salesmanship. Chatbots, or at least the way the industry designs them, are inherently addictive and engaging; a Google search is not.
In any case, OpenAI wouldn’t be alone in going down this route. Google has been showing sponsored content in its infamously wonky AI Overviews for over a year. Perplexity has been experimenting with ads since last year, too, as has the Google-backed AI companion platform Chai.
Je pense que comme moi, vous n’avez jamais vu un fauteuil roulant monter des escaliers ? Hé ça va changer en lisant cet article puisque Toyota vient de présenter un truc complètement dingue lors du Japan Mobility Show 2025. Il s’agit d’un fauteuil roulant autonome avec des pattes articulées qui lui permettent de se déplacer comme un crabe.
Le bestiau s’appelle “Walk Me” et c’est pas juste un concept de designer sous acide puisque ce machin dispose de quatre pattes repliables inspirées de la locomotion animale, notamment celle des chèvres, des crabes et des petits vieux chez Auchan le mardi matin à 8h30. Et quand je dis inspirées, c’est pas du bullshit marketing car les ingénieurs de Toyota ont vraiment étudié comment ces animaux se déplaçaient sur des terrains accidentés pour reproduire leurs mouvements.
Concrètement, quand le Walk Me grimpe un escalier, les papattes avant déterminent la hauteur de la marche et tirent la structure vers le haut, pendant que les papattes arrière maintiennent l’équilibre et poussent le corps. C’est fluide, c’est stable, et surtout, contrairement à vous qui ne pouvez rien faire sans votre femme, c’est autonome ! Donc pas besoin de quelqu’un derrière pour vous porter ou vous pousser.
Et pour éviter de se vautrer ou de percuter des obstacles, le Walk Me embarque des capteurs LiDAR et des caméras qui scannent en permanence l’environnement. Le système crée alors une représentation 3D en temps réel de ce qui l’entoure et s’adapte automatiquement comme ça, si un obstacle apparaît (genre un bébé ninja), il s’arrête tout seul !
Et le truc malin, c’est que quand vous n’avez pas besoin des pattes, elles se replient sous le siège, du coup le fauteuil devient suffisamment compact pour rentrer dans un coffre de voiture ou dans vos bagages. Le système peut même se déployer et se stabiliser tout seul sans intervention de l’utilisateur.
Bon, bien sûr, le Walk Me reste un prototype sans date de commercialisation annoncée mais sa présentation au Japan Mobility Show montre que Toyota prend au sérieux la mobilité inclusive. Y’a encore du boulot avant de voir ce genre d’engin dans les rues, mais ça donne quand même un bon aperçu d’un futur où les fauteuils roulants ne seront plus limités par les escaliers, les trottoirs défoncés ou les terrains un peu roots qu’on ne connait que trop bien en France…
The first step was to make the holographic segment displays, because they’re not really something you can just buy off the shelf. [mosivers] achieved this by using a kit from LitiHolo, which enables you to create holograms by shooting a laser at special holographic film. Only, a few upgrades were made to use the kit with a nicer red diode laser that [mosivers] had on hand for better performance. The seven-segment layouts were carefully recorded on to the film to form the basic numerals of the clock, such that illuminating the films from different angles would light different segments of the numeral. It’s quite involved, but it’s explained well in the build video.
As for the timekeeping side of things, an ESP32 was used, setup to query a network time server to stay accurate. The microcontroller then commands a series of LEDs to light up as needed to illuminate the relevant segments of the holographic film to show the time.
Ultimately, [mosivers] built a cool clock with a look you won’t find anywhere else. It’s a lot more work than just wiring up some classic seven-segment LEDs, but we think the result is worth it. If you fancy other weird seven-segment builds, though, we’ve got plenty of others in the till.
Closed-cell self-expanding foam (spray foam) is an amazing material that sees common use in construction. But one application that we hadn’t heard of before was using it to fill the internal voids of 3D printed objects. As argued by [Alex] in a half-baked-research YouTube video, this foam could be very helpful with making sure that printed boats keep floating and water stays out of sensitive electronic bits.
It’s pretty common knowledge by now that 3D printed objects from FDM printers aren’t really watertight. Due to the way that these printers work, there’s plenty of opportunity for small gaps and voids between layers to permit moisture to seep through. This is where the use of this self-expanding foam comes into play, as it’s guaranteed to be watertight. In addition, [Alex] also tests how this affects the strength of the print and using its insulating properties.
The test prints are designed with the requisite port through which the spray foam is injected as well as pressure relief holes. After a 24 hour curing period the excess foam is trimmed. Early testing showed that in order for the foam to cure well inside the part, it needed to be first flushed with water to provide the moisture necessary for the chemical reaction. It’s also essential to have sufficient pressure relief holes, especially for the larger parts, as the expanding foam can cause structural failure.
As for the results, in terms of waterproofing there was some water absorption, likely in the PETG part. But after 28 hours of submerging none of the sample cubes filled up with water. The samples did not get any stronger tensile-wise, but the compression test showed a 25 – 70% increase in resistance to buckling, which is quite significant.
Finally, after tossing some ice cubes into a plain FDM printed box and one filled with foam, it took less than six hours for the ice to melt, compared to the spray foam insulated box which took just under eight hours.
This seems to suggest that adding some of this self-expanding foam to your 3D printed part makes a lot of sense if you want to keep water out, add more compressive strength, or would like to add thermal insulation beyond what FDM infill patterns can provide.
Après avoir démontré que les souvenirs s’inscrivent de façon matérielle dans notre cerveau, les scientifiques s’attèlent à la tâche de réactiver des souvenirs attaqués par des maladies neurodégénératives chez des souris.
Le monde numérique évolue à une vitesse fulgurante, mais les compétences informatiques des jeunes générations semblent régresser. Des professeurs tirent la sonnette d'alarme face à des élèves incapables d'effectuer des tâches basiques sur ordinateur. Cette situation soulève des questions...
Des chercheurs britanniques ont travaillé pendant dix années sur une hypothèse autour des superbactéries. Ils ont été stupéfaits lorsque l’IA de Google, Co-scientist, est arrivée à la même hypothèse en seulement 48 heures.
"If you have an apple and I have an apple and we exchange apples, then you and I will still each have one apple. But if you have an idea and I have an idea and we exchange these ideas, then each of us will have two ideas."
This quote, commonly attributed to George Bernard Shaw, though probably not from him, highlights one of the lovely things about ideas. They behave differently from physical things. They're not exclusive; they're additive and abundant.
Ideas Don't Behave the Same as Apples
Though we talk about intellectual property, ideas don't behave like property in the usual sense. One of the simplest ways to see the difference between ideas and objects is to look at what happens when we share them.
I can give you an idea, and we both have the idea, but if I give you my apple, then I no longer have one. This also makes ideas very hard to take back once they are out. Ideas are harder to control than objects.
Because ideas are abstract—they don't exist in a physical form—we use conceptual metaphor to talk and reason about them. What follows are some of my favourite examples of how we think about ideas, drawn primarily from Philosophy in the Flesh and the very readable Metaphors We Live By by George Lakoff and Mark Johnson.
The Ideas Are Objects Metaphor
Metaphors are how we talk about abstract concepts, like an idea. And the Ideas Are Objects metaphor is one of the most common ways we understand them. This means that a lot of how we talk about apples and ideas overlap.
So I can:
give you an idea (or an apple), and maybe you'll
grasp it, or perhaps it might
go over your head
Ideas can be:
solid or weak
shared or hoarded
fragile or bulletproof
cheap or gold
Thinking Is Object Manipulation
Because you can examine objects and manipulate them, we have the metaphor Thinking Is Object Manipulation.
When you think about an idea, you might:
play with it
toss it around
see if it sticks
share it with others
try to break it
When you communicate with someone, you exchange ideas.
you can give an idea
get an idea across
Together with the Mind Is a Body metaphor (mental exercise), a teacher might try to:
put an idea into students' minds
fill them up during the term
see how much they've retained
And I've certainly employed cramming before a test.
Because Understanding Is Grasping—having an idea under control—when you don't understand something, it might:
be slippery
resist definition—has no shape
be beyond your grasp
As Thinking Is Object Manipulation, you can work on an idea:
reshape it
craft it
fashion it
analyse it by taking it apart
deconstruct it
Together with Knowing Is Seeing, we can:
Turn an idea over to see both sides of it
hold it up to scrutiny
shine a light on it
put it under the microscope
Ideas Are Food and Acquiring Ideas Is Eating
Another fun and common metaphor for ideas is rather more like apples. If the Mind Is a Body, then we need to feed it healthy, nutritious food. I like to think readers of Sketchplanations have an insatiable curiosity. In much of our reasoning, then, ideas are a special kind of object—they are our food for thought.
Unhelpful ideas are unhealthy, and helpful ideas are healthy, so they might be:
raw
fresh
half-baked
sweet
Or an idea might be:
rotten
disgusting or unsavory
unpalatable
hard to digest
Or they might:
smell fishy
leave a bad taste in your mouth
Or perhaps they need to be:
put on the back burner
chewed on for a while
sugar-coated
Significant ideas are:
meaty
something to chew on
let stew for a while
So What Are Ideas Really Like?
So, given all this talk about Ideas As Objects, it's easy to assume that ideas should behave like them. But before they take on a physical form — say, as a building or product — they don't behave the same as objects.
Hopefully, this sketch conveys this idea to you while keeping it with me, too.
I think a more suitable conceptualisation is the magical powers of software with its infinite copy/paste. As software becomes more of our daily experience, we might gradually adopt more accurate metaphors for how ideas actually behave.
These are not the only metaphors we use across languages for understanding and reasoning about ideas and thinking, Ideas Are Locations, and Thinking Is Moving in particular ("I don't follow you"). There are many. Do check out Metaphors We Live By if you're curious for more (or Philosophy in the Flesh if you really want to get stuck in).
