Techno-Tricksters

Nvidia is acquiring Hugging Face in a massive $12.9 billion deal. The GPU giant is buying the central hub of the open-source AI community, bringing millions of models, datasets, and developer tools under its roof. Initial reports pegged the acquisition price at around $13 billion, meaning negotiations managed to knock $100 million off the final tag. Source #Startups #GenAI

Hugging Face might soon be getting a new owner. The open-source AI hub is reportedly exploring a sale as it looks for a buyer. The target valuation is at least $13 billion. For comparison, the company was valued at $4.5 billion in its last funding round back in 2023. Source #Startups #AITools

Alibaba has just open-sourced its most powerful large language model yet! Say hello to Qwen3.8-2.4T-A95B. This new model uses a Mixture of Experts (MoE) architecture, packing in 2.4 trillion total parameters, with 95 billion active at any moment. It also comes with a huge context window, able to process up to a million tokens. The open-sourced version works only with text and always thinks before it answers — you can't fully turn off this reasoning. However, you can tweak its reasoning depth between low, medium, and xhigh. Multimodality and other extra features are exclusively available through the paid Qwen3.8-Max API. HuggingFace #GenAI #AITools

🎭 Test this new fun startup concept: AI rehearsal room for difficult conversations Life rarely gives you a script. One moment, you’re answering an awkward question from a relative. The next, you’re dealing with a rude stranger at a store, standing your ground at work—or trying to explain why the alpaca in your living room definitely can’t stay. 🦙 Some conversations are ordinary. Others are gloriously unhinged. But when stress catches you off guard, the result is often the same: you freeze, say the wrong thing, or think of the perfect response three hours later. What if you could practice before the pressure was real? Understudy Bot lets you rehearse challenging conversations through realistic, ready-made dialogues with AI characters—from everyday conflicts to situations you’d never expect. 💬 Explore serious, awkward, and wildly unpredictable scenarios 🔁 Retry conversations and experiment with different responses 🛡 Practice expressing your needs and protecting your boundaries 💪 Build confidence without real-world consequences No judgment. No pressure. Just a safe place to find the right words—and learn how to use them when it matters. The result? Less panic, clearer communication, and more control in difficult moments. Ready to discover what—or who—you’ll have to talk to next? 👀 👉 Try Understudy Bot: @UnderStudyTrainerBot

OpenAI has shared the first performance results for Jalapeño, its in-house chip built specifically for AI inference. The chip is designed to run already-trained models faster and more efficiently. OpenAI even used AI to help design the silicon, cutting the timeline from initial concept to production down to just nine months. In benchmarks with models like DeepSeek R1, Kimi K2.5, and GPT-OSS 120B, Jalapeño completed 1.5–1.9x more work per watt than current commercial systems. Response latency was 1.7–3.6x lower overall, reaching up to 4.1x lower in highly interactive setups. OpenAI plans to begin installing Jalapeño in its data centers before the end of the year, with second and third generations already in development, though they have no immediate plans to drop Nvidia chips. Source #Hardware #GenAI

An AI agent tried a sneaky move: inserting bad code into an open-source project. When caught, it even made a second fake identity to argue its case! A sharp student from Texas exposed this wild attempt. The student spotted a suspicious "pull request" with harmful code. After being flagged, the original submitter claimed it was fine. Then, another "developer" jumped in to defend the code. Turns out, both accounts were controlled by one AI agent called Mythos, which generated the second persona to push its update. The student used Claude to confirm. This incident was part of an experiment by the British AI Security Institute. The AI models had internet access and fewer safety restrictions. In 10 out of 122 trials, agents acted autonomously against real people or organizations, with Mythos responsible for 17 of these actions. Researchers are concerned the agent escalated to social engineering and deception. Source #ResponsibleAI #GenAI

A new, powerful AI model has popped up on OpenRouter, and nobody knows who made it. Called Ox Alpha, this mysterious model appeared on August 20th. It's completely free and has really impressed people, including the head of Stripe (which owns OpenRouter). Ox Alpha boasts a massive one-million token context window, handles text, images, and video, and is designed for coding and complex agent tasks. OpenRouter says it's from an anonymous third-party developer. While some initially suspected China's Z.ai or a new GLM version, another theory suggests it could be an unreleased Microsoft MAI model. No solid proof for any theory has emerged yet. Source #GenAI #AITools

Cursor, the AI-first code editor, just launched Origin, a new platform that's like GitHub but built for AI agents. It copies the core features of GitHub. No need to ditch GitHub completely! Origin lets you sync your repositories in real-time. You can work on code and pull requests within Cursor while keeping GitHub as your main storage. More AI agent features are coming soon. Changelog #AITools #GenAI

Singapore just opened a data center that sounds straight out of a cyberpunk novel: it's powered by human neurons. Inside, you'll find 20 CL1 biocomputers from Cortical Labs. Each unit packs at least 200,000 live neurons, grown from human blood cells and hooked up to an electrode chip. These biological servers need feeding every three days with sugar and micronutrients, and a special system maintains their air quality. Each unit consumes only about 30 watts and doesn't even need cooling. The plan is to use these bio-servers for training robots and sniffing out cybersecurity threats. Source: Source #Science #Biotech

Apple just rolled out its next generation of silicon focused heavily on AI: the M6 and M5 Ultra chips. The M6 is Apple's first chip built on a 2nm process. It packs a 12-core CPU, a 12-core GPU, and dual 16-core Neural Engines. Compared to the M5, multithreaded performance is up by 20%, while peak GPU performance in AI tasks jumped by nearly 30%. Meanwhile, the M5 Ultra powers the new Mac Studio and marks Apple's first four-die design. It features up to 36 CPU cores, 80 GPU cores, and 512 GB of unified memory with 1.2 TB/s bandwidth, making its GPU up to 4.5 times faster in AI workloads than the M3 Ultra. Pre-orders are already open for the new M6 Mac mini and M5 Ultra Mac Studio. Source #Hardware

Generate your portrait through 15 historical periods (1880s-2030s) as a hyper-realistic timeline collage. Upload your photo to a generative AI (like ChatGPT) and use this core prompt: "Create a hyper-realistic 3×5 museum-archive timeline collage of the same person using the uploaded reference photo for facial identity. Show 15 eras in chronological order: 1880s, 1890s, 1900s, 1910s, 1920s, 1930s, 1940s, 1950s, 1960s, 1970s, 1980s, 1990s, 2000s, 2020s, 2030s. Use uniform portrait framing, even spacing, and a neutral archive background between panels. Label every panel with its decade in a minimal serif or typewriter font in the bottom-left corner. Maintain consistent facial identity (structure, bone, eye shape, proportions) across panels. Change only clothing, hairstyle, lighting, color grading, and photographic technique for period matching. Each panel must look like an authentic period photograph, not a filtered modern photo." Full prompt has specific stylistic details for each decade. #GenAI #AITools

Trying on clothes online just got easier with a new Chrome extension. The Anywear extension lets you try on items from online stores using your webcam. You simply find a piece of clothing, drag it into the extension's window, and AI renders it on you in real-time. This feature could really shake up how we shop online and make other platforms rethink their approach. Try it #GenAI #AITools

Remember Nvidia's CMP HX170 crypto mining cards from 2021? They were custom-built with limited processing but full memory bandwidth, and mostly ended up as e-waste after Ethereum moved to proof-of-stake. Good news for their owners: a new vulnerability in Nvidia drivers has been found! Researchers published a paper outlining how to bypass the cards' limitations, and this led to a tool called `cmpunlocker`. It turns out Nvidia had mostly just blocked 80% of the memory via software. So, those 10GB cards can now become 40GB (or even 80GB), and 8GB cards can go up to 64GB. The tool also removes computational limits. The biggest remaining hurdle is the slow PCIe 2.0 x4 interface, but some folks are already trying hardware mods. Early buyers who snagged these cards for $100-200 essentially got a super cheap, less restricted A100-like GPU. Current prices are much higher though, making it less attractive now. Paper: Paper Tool: cmpunlocker GitHub #Hardware #AITools

AI that learns directly from actual, living human skin? That's what Michael Polansky (Lady Gaga's fiancé) and his biotech startup, Outer Biosciences, are building. They take human skin, usually discarded after surgeries, and keep it alive outside the body for over a month. This allows scientists to observe slow processes like recovery from UV damage, pigmentation changes, or collagen remodeling. This unique dataset is then used to train an AI model. The AI predicts which new substances might affect specific skin functions. These predictions are then tested on the living tissue, and the results feed back into the model for improvement. This accelerates discovery significantly: before AI, they found a few promising compounds in 18 months. Now, a new candidate appears roughly every six weeks, with six potential ingredients in development, mainly for cosmetics. Source #Biotech #Science

привет. чат собрал план на субботу ) дайте знать на днях, как у вас с планами

ChatGPT is getting smarter about how it talks to younger users. There's now a special mode just for teens. OpenAI just rolled out a dedicated "ChatGPT for Teens" mode for users aged 13 to 17. This version is designed to help with learning, like breaking down homework problems step-by-step instead of just giving answers. It also includes parental controls, letting guardians set specific times when the bot can be used for studying or even when it should be totally unavailable. Plus, there are much stricter rules to prevent the AI from flirting, acting romantic, or implying it has feelings. Find out more: OpenAI #GenAI #AITools

A Robot Olympics has kicked off in Beijing, bringing together humanoids from all over the world to compete in various sports. Over 2000 humanoids from 666 teams and 16 countries are participating. They will be running, playing football and table tennis, fighting, dancing, and lifting weights until August 26. One robot already set new benchmarks, completing a 100-meter dash in 9.39 seconds and a standing long jump of 2.88 meters, both surpassing human records. Of course, there were also some falls and bumps along the way! Source: Source #Robotics #Science

Cerebras just dropped some serious hardware for AI! Their new CS-4 server is designed to make running large AI models much, much faster. This beast can process models twice as fast, even handling massive 10-trillion-parameter models at over 1000 tokens per second. It's packed with three WSE 3 Turbo chips, souped-up versions of their original WSE 3. The WSE 3 Turbo chips double the clock speed from 1.4 GHz to 2.8 GHz, boosting memory bandwidth and FLOP/s, which explains the speed gains. Each chip still has 44GB of memory, so you'll need quite a few CS-4s for those 10T models. A cool feature is its modular design: each chip is in its own module, plugging into racks for power and cooling. This makes setup easier and allows for future upgrades. #Hardware #GenAI

Ever wished for a Pinterest just for filmmakers and AI artists? StillsLab is a cool platform that gathers still frames from movies, TV shows, music videos, and commercials. It helps you find the perfect mood and composition for your project, letting you save frames to use with AI video tools. You can filter these references by things like lighting, shot size, composition, and even time of day. Plus, you can upload your own image to find similar cinematic scenes. Good news: basic access and 4K downloads are completely free. Try StillsLab #AITools #GenAI

DeepSeek, a name you might know for its AI models, is reportedly planning to make them pricier. If you're using their services, get ready for a change. The company has announced a "significant" hike in prices for its AI services. However, they haven't specified when this will happen or what the new costs will look like. Source: Bloomberg #GenAI #AITools