What just happened? Microsoft has introduced BitNet b1.58 2B4T, a new type of large language model engineered for exceptional efficiency. Unlike conventional AI models that rely on 16- or 32-bit floating-point numbers to represent each weight, BitNet uses only three discrete values: -1, 0, or +1. This approach, known as ternary quantization, allows each weight to be stored in just 1.58 bits. The result is a model that dramatically reduces memory usage and can run far more easily on standard hardware, without requiring the high-end GPUs typically needed for large-scale AI. The BitNet b1.58 2B4T model was developed by Microsoft's General Artificial Intelligence group and contains two billion parameters – internal values that enable the model to understand and generate language. To compensate for its low-precision weights, the model was trained on a massive dataset of four trillion tokens, roughly equivalent to the contents of 33 million books. This extensive training allows BitNet to perform on par with – or in some cases, better than – other leading models of similar size, such as Meta's Llama 3.2 1B, Google's Gemma 3 1B, and Alibaba's Qwen 2.5 1.5B. In benchmark tests, BitNet b1.58 2B4T demonstrated strong performance across a variety of tasks, including grade-school math problems and questions requiring common sense reasoning. In certain evaluations, it even outperformed its competitors. What truly sets BitNet apart is its memory efficiency. The model requires just 400MB of memory, less than a third of what comparable models typically need. As a result, it can run smoothly on standard CPUs, including Apple's M2 chip, without relying on high-end GPUs or specialized AI hardware. This level of efficiency is made possible by a custom software framework called bitnet.cpp, which is optimized to take full advantage of the model's ternary weights. The framework ensures fast and lightweight performance on everyday computing devices. Standard AI libraries like Hugging Face's Transformers don't offer the same performance advantages as BitNet b1.58 2B4T, making the use of the custom bitnet.cpp framework essential. Available on GitHub, the framework is currently optimized for CPUs, but support for other processor types is planned in future updates. The idea of reducing model precision to save memory isn't new as researchers have long explored model compression. However, most past attempts involved converting full-precision models after training, often at the cost of accuracy. BitNet b1.58 2B4T takes a different approach: it is trained from the ground up using only three weight values (-1, 0, and +1). This allows it to avoid many of the performance losses seen in earlier methods. // Related Stories This shift has significant implications. Running large AI models typically demands powerful hardware and considerable energy, factors that drive up costs and environmental impact. Because BitNet relies on extremely simple computations – mostly additions instead of multiplications – it consumes far less energy. Microsoft researchers estimate it uses 85 to 96 percent less energy than comparable full-precision models. This could open the door to running advanced AI directly on personal devices, without the need for cloud-based supercomputers. That said, BitNet b1.58 2B4T does have some limitations. It currently supports only specific hardware and requires the custom bitnet.cpp framework. Its context window – the amount of text it can process at once – is smaller than that of the most advanced models. Researchers are still investigating why the model performs so well with such a simplified architecture. Future work aims to expand its capabilities, including support for more languages and longer text inputs.

For being such an innovative narrative adventure game, Lost Records: Bloom & Rage still adheres to a release format that no longer works for the genre. Lost Records is one of the best games released in 2025 so far. It tells a timeless coming-of-age tale that isn’t afraid to revel in 90s nostalgia. It uses the choice-driven adventure game format trailblazed by Telltale and Don’t Nod’s early work, but infuses it with clever ideas like splitting perspectives between the past and future and letting players film the world with a gyroscopic-controlled camcorder. I highly recommend you go and play it now that it’s all out, but I feel awkward that I couldn’t wholeheartedly recommend it until now. Recommended Videos Dont’ Nod decided to release Lost Records in two parts. Tape 1 launched on February 18, while Tape 2 came out on April 15. While Don’t Nod hoped this approach would give players time to stop and reflect on the game, it had the opposite effect. In an era where so many games are coming out and simultaneously competing for attention, the episodic release isn’t nearly as alluring as it was in 2014, especially if it’s only in part. Lost Records: Bloom & Rage | LAUNCH TRAILER - Tape 2 Lost Records follows a group of friends over the summer of 1995. They’re a group of misfits who come together and create an inseparable bond, especially when they discover a hidden outcove and a glowing supernatural hole in the wilderness near their hometown. Tape 1 is mainly about the friend group meeting up and forming the titular band, Bloom & Rage. It culminates with the band’s first performance and a revelation that recontextualizes some of what has been going on. Related Tape 2 follows the group as they deal with the fallout of the event of the first game. Alongside all of this, Lost Records also follows three friends reuniting in the modern day to reflect on their past and open a mysterious box left for them. It’s a compelling story when all is said and done, but not one I think had to be chopped into two parts. The slower pace of the narrative makes the individual halves feel unfulfilling on their own. Narratively, many of the main themes and mysteries don’t coalesce until you play the whole thing. Don’t Nod has given its reasoning for releasing the game in two parts. Creative Director Michel Koch told Game Rant that he likes “media that values my time and doesn’t try to lock me into endless hours of content.” To Eurogamer, Koch elaborated that he enjoys weekly series on TV and that, after the ending of Tape 1, it made sense for Don’t Nod to give players space to “stop there and reflect a bit” so they can come back to Tape 2 “even more invested in the story because you have had some time to reflect on what happen.” Short games are just as important as 100+ hour live service experiences, and weekly TV has its charms. While I agree with Koch on both of those points, Tape 1 and Tape 2 of Lost Records aren’t structured like television episodes and the whole game only takes about 8 or 9 hours to complete. Even if there are aspects of its story to reflect on, having a two-month wait between two four-hour chunks of a cohesive story didn’t benefit Lost Records at large. If the goal was to emulate TV, then Lost Records‘ narrative structure could’ve been split up into a more TV episode-like structure. Don't Nod While I enjoyed Tape 1 of Lost Records, I might not have had the time to play Tape 2 had I not planned to write about it, because it’s coming out in a shockingly crowded month for games. I also couldn’t give it a wholehearted recommendation until now because it was an incomplete experience. Because so many games compete for players’ time and attention nowadays, it’s a risk to ask players to invest in the first half of your adventure game and then make them wait two months to finish the story. Looking at the limited data we have, the Tape 2 launch on Steam has accrued higher player counts than Tape 1’s launch did. Would the numbers have been stronger from release had the whole game come out simultaneously? I like Lost Records and hope lots of people play it, but part of me is retrospectively looking back and wondering if the episodic release model was a good fit for this new Don’t Nod narrative adventure. Thankfully, we no longer have to worry about that as all of Lost Records: Bloom & Rage is now available across PC, PS5, and Xbox Series X/S. It’s part of the PS Plus game catalog on PS5. Editors’ Recommendations

Paulo Sousa/Getty Images 2025-04-20T16:45:02Z Save Saved Read in app This story is available exclusively to Business Insider subscribers. Become an Insider and start reading now. Have an account? I am a therapist and mom of four. I spent months trying to "fix" my second child's emotional outbursts and meltdowns. Learning to work with sensitivity instead of fighting it changed my parenting approach. "I have good news! No big outbursts today or crying on the floor; he just needed some extra support during our transition times!"The sunny, early afternoon light streamed through Ms. Barb's preschool classroom window as she delivered what was meant to be encouraging news. But her description of my son's day hurt. How was it that a "good day" for him was not having a full-blown tantrum in the middle of class?As I drove him home, squinting against the bright sunshine, he napped quietly in the backseat. I gripped the steering wheel tighter, fighting back tears. I felt like I was messing everything up. Out of our four kids, his behavior felt uncontrollable, and I was constantly on the search for ways to "fix" him.The late-night search that changed everythingEvery night after tucking the kids in, I would end up in my usual spot on our couch, scrolling through parenting articles. My eyes would be heavy, but my mind would be racing. What could I be doing differently? That's when I saw an article about highly sensitive children, and I stopped scrolling.I sat up straighter as I read each line. These kids hate crowds. They can't stand loud noises. They pick up on everything around them, are highly intelligent, and feel emotions like they're turned up to full volume. The author learned about parenting a highly sensitive person Courtesy of the author This explained why he would cover his ears and cry when things got too chaotic around him. No wonder he melted down at preschool, church, and birthday parties — it was because he was overstimulated. Everything I'd been beating myself up about now made perfect sense. But there was something else. This list wasn't just describing my son, it was describing me too!Finding myself in my child's storyAs I read more about highly sensitive people (HSP), memories from my own childhood came rushing back to mind. At 14, I remember sitting in my bedroom after school, overwhelmed by my friend's breakup — feeling her pain as intensely as if it were my own. I called it "over-empathizing" back then, and it eventually led me to become a therapist. But until this moment, I never understood why I felt everything so much more deeply than others seemed to. It also explained why I pick up on tiny details that others miss and why I get easily overwhelmed. I realized in this moment that all this time, I had been trying to fix my son when, really, he was helping me understand a part of myself that had always felt like a flaw.A new way of parenting and livingThe discovery that night changed so much about how I live and parent now. I stopped seeing my son's sensitivity as something that needed to change. Instead, I saw it as a trait that was just a part of him and that we could work with.We now leave birthday parties a little early if he starts to get overwhelmed. We make sure that he has downtime after school to decompress. He and I talk about being HSP, which helps him process his big emotions.And just like him, I started to figure out how to schedule my life in a way that works with, not against my sensitivity. I no longer feel guilty about putting on noise-canceling headphones when my kids get too loud and overstimulating.Coming full circleWhat started as a desperate attempt to fix my son's tantrums turned into one of the best things that has happened in my life. It's finally having the permission to be exactly who I am and allowing my son to be exactly who he is. Now, when parents come to my therapy practice describing their "difficult" or "overly emotional" child, I see the familiar signs. I share my story, watching relief wash over their faces as they begin to understand their child differently.Last week, a mom teared up when I explained how her daughter's sensitivity wasn't a problem to fix but simply part of who she is. These sessions have shifted from problem-solving missions into conversations where parents discover new ways to work with their child's sensitivity.Because oftentimes, the very things that we are trying to change about our children are the things that can teach us the most about ourselves. Recommended video

All weekend at Star Wars Celebration Japan, the connection between the galaxy far, far away and its inspirational roots in Japanese media has been everywhere. But perhaps one of the most joyful celebrations of that connection was tucked away among the hustle and bustle of the show floor. At the back side of a special anime and manga pavilion on the Celebration show floor, primarily dedicated to displays of concept art and props from the first two volumes of Star Wars Visions, a whole wall of art invited congoers to see a special mashup gallery merging the worlds of Studio Ghibli and Star Wars. Featuring art from a bevy of talent from across Lucasfilm divisions—spearheaded by, of course, none other than noted Ghibli fan Dave Filoni himself—the gallery saw characters and ideas from across Star Wars merge with iconic Ghibli films and imagery, from Porco Rosso to Princess Mononoke, and even touching on Lucasfilm’s actual official collaboration with the studio, Zen: Grogu and Dust Bunnies. “Japan has always influenced my creativity. I grew up loving Tokusatsu, and was captivated by anime series like Macross,” Filoni wrote in a message accompanying his own art piece in the collection, “Ahsokanoke”, depicting Ahsoka and Grogu in a Ghibli-esque style. “After I saw Miyazaki’s film Princess Mononoke, it changed the way I thought about animation. I loved everything about that film and it made a lasting impression on me. When I went to create Ahsoka Tano, I knew I wanted to imbue her with a bit of San. So this is why Ahsoka has two ‘wolf fangs’ on her cheeks, in homage to Miyazaki’s film.” Check out a rundown of the Lucasfilm x Ghibli gallery below. Star Wars Celebration Japan Lucasfilm Ghibli Mashup Art Gallery © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 © James Whitbrook/io9 Want more io9 news? Check out when to expect the latest Marvel, Star Wars, and Star Trek releases, what’s next for the DC Universe on film and TV, and everything you need to know about the future of Doctor Who.

Nature, Published online: 16 April 2025; doi:10.1038/s41586-025-08845-yBilateral grafts of cryopreserved human embryonic stem cell-derived dopaminergic neuron progenitor cells into the putamen of patients with Parkinson’s disease in a phase I clinical trial showed safety, improvements in off-drug motor function and graft survival at 18 months after transplant.

A decorated duck egg discovered in a trash pit is the world's oldest "pysanka."

RT Creative ShrimpHave you heard about the TrueDepth addon that brings the DepthAnything model (monocular depth estimation) into Blender? We've just tested it:

The April #NVIDIAStudio Driver is here. ✅Download now for the latest optimizations ➡️ https://nvda.ws/42x5JTn

MCP vs Function Calling : Comment les LLM interagissent avec le monde réel2 min read·Just now--IntroductionDepuis l’arrivée des modèles de langage avancés (LLM) comme GPT-4, Claude 3 ou DeepSeek, une question essentielle se pose : comment ces modèles peuvent-ils interagir efficacement avec des outils, des APIs ou des systèmes externes ?Deux grandes approches se sont imposées :MCP (Multi-step Conversational Prompting)Function CallingDans cet article, je vous explique ces concepts, qui les utilise aujourd’hui et dans quels cas les appliquer.MCP : Multi-step Conversational PromptingLe MCP est une méthode où le LLM agit comme un agent conversationnel intelligent qui procède par étapes successives pour traiter une demande.Comment ça fonctionne ?1. Le LLM analyse la demande.2. Il propose des options ou demande des précisions.3. Il choisit dynamiquement l’outil ou l’API à utiliser.4. Il exécute l’action et restitue la réponse.Pourquoi c’est utile ?Contrôle conversationnel : idéal dans des situations sensibles (santé, finance).Gestion multi-API : il peut gérer plusieurs services dans un seul échange.Workflows complexes : permet de gérer des dialogues à plusieurs étapes.ExemplesAssistant bancaire : validation étape par étape avant un virement.Diagnostic médical : questions complémentaires avant une recommandation.Service client : suivi de commande + demande de retour dans la même conversation.Function CallingLe Function Calling est une capacité intégrée dans certains modèles où le LLM détecte automatiquement qu’il doit appeler une fonction externe pour répondre à la demande de l’utilisateur.Comment ça fonctionne ?1. L’utilisateur formule une demande.2. Le LLM génère un appel structuré (nom de fonction + arguments).3. Le système exécute la fonction/API.4. Le résultat est restitué par le modèle.Pourquoi c’est utile ?Rapidité : pas besoin de multi-étapes.Automatisation : déclenche directement des actions concrètes.Simplicité : idéal pour des cas d’usage bien cadrés.ExemplesApplication météo : « Quel temps fait-il à Paris ? »Convertisseur de devises : « Convertis 100 dollars en euros. »Réservation : « Réserve-moi une table ce soir à 20h. »Qui les utilise aujourd’hui ?MCPAzure OpenAI CopilotLangChainSalesforce Einstein CopilotDeepSeek (via LangChain ou agents personnalisés)Function CallingOpenAI GPT-4 et GPT-3.5 TurboAnthropic Claude 3 (Tool Use)Google GeminiLlama 2 et Mistral (via LangChain)DeepSeek (possible via orchestrateurs, non natif)pComparatifConclusionMCP et Function Calling sont deux approches complémentaires.MCP est parfait pour les situations complexes et sensibles où chaque étape compte.Function Calling brille dans les cas rapides et automatisables.Avec l’évolution des modèles comme GPT-4, Claude 3, Gemini ou DeepSeek, ces méthodes permettent de créer des assistants plus intelligents et réellement connectés à des services externes.