NVIDIA's PiD is a new pixel diffusion decoder for... high-res image models. It skips decode-then-upscale stop, making sharper outputs faster. > Directly generates 4K images > Up to 5.9x faster than SeedVR2 > Free & open weightsshow more

> be fal > release new video model >... generates 15 sec video > takes 3.5 sec > video literally takes longer to watch than generate > somehow faster than image models > user is now the latencyshow more

P-Image-Upscale is the fastest and cheapest image upscaler in... the world: supporting outputs up to 128 MP under 1 seconds. A couple of weeks ago we released p-image-upscale, and it now works better than ever. It’s the fastest image upscaler in the world, supporting outputs up to 128 MP while keeping pricing simple and predictable: - $0.005/image for 1–4 MP - $0.01/image for 4–8 MP - $0.02/image for 8–16 MP - $0.04/image for 16–32 MP - $0.06/image for 32–64 MP - $0.12/image for 64–128 MP That means you can go from low-res input to production-ready output with extreme speed, while preserving detail and keeping costs easy to understand. Available on: - Pruna AI | - each::labs | - inference shell | - Replicate | - Runware | - Segmind | - WaveSpeedAI | - Wiro AI — Ship AI Faster | If you’re building workflows where image quality, price, latency, and scale all matter, p-image-upscale is built for you.show more

Bytedance just dropped realtime voice translation 3x faster than... before, with only a ~3s lag! Seed LiveInterp 2 is a full duplex speech-to-speech model with >70% correctness. When this makes it to video calls, it'll open up previously impossible connections.show more

🌟 New mini feature today on Photo AI It... now auto recognizes your model's attributes when you upload photos of yourself Like gender, age, ethnicity, eye color, etc This uses xAI's Grok because it's the least restricted and other models don't let you do any identity type stuff due to safety (???) This saves new users a few more seconds when creating new models: training a model takes only about 30 seconds now, and then it also auto generates free sample pics for you in about 5-10 seconds So in under a minute of signing up and uploading a few selfies you got your first AI photo shoot pics I learnt the faster I get new users to the classic "WOW" moment the happier they get and the less churn I get 😊show more

🚨 JUST IN: THIS FREE TOOL JUST REPLACED FOUR... AI IMAGE AND VIDEO SUBSCRIPTIONS AT ONCE. Midjourney. Krea. Higgsfield. Openart. One repo. 200+ models. Zero dollars a month. Here is what it actually does. It is a full image and video studio that runs in your browser or as a desktop app. Text to image, image to image, text to video, image to video, lip sync, cinema mode with real camera controls. All of it. 4,500 people already starred this. What you get for free: → 50+ image models including Flux, Midjourney v7, Ideogram, GPT-4o, Seedream → 60+ video models including Kling, Sora, Veo, Runway, Wan, Hailuo → lip sync studio with 9 dedicated models. upload a portrait and audio and it talks → cinema studio with real camera controls. lens, focal length, aperture, film stock → feed up to 14 reference images into one generation → self-hosted. your data never leaves your machine The crazy part is there is also a hosted version that needs zero setup. Just open the link and start generating. Now the math. Midjourney Standard: $30/month Krea AI Pro: $30/month Higgsfield Plus: $49/month Openart AI: $15/month That is $124 a month. $1,488 a year. This repo does everything all four do. With more models than any of them. For free. Forever. No subscription. No vendor lock-in. MIT licensed. Download it in one click on Mac or Windows. Someone should have told me about this sooner. I feel like an idiot. ( save this )show more

STEVE-1: A Generative Model for Text-to-Behavior in Minecraft paper... page: Constructing AI models that respond to text instructions is challenging, especially for sequential decision-making tasks. This work introduces an instruction-tuned Video Pretraining (VPT) model for Minecraft called STEVE-1, demonstrating that the unCLIP approach, utilized in DALL-E 2, is also effective for creating instruction-following sequential decision-making agents. STEVE-1 is trained in two steps: adapting the pretrained VPT model to follow commands in MineCLIP's latent space, then training a prior to predict latent codes from text. This allows us to finetune VPT through self-supervised behavioral cloning and hindsight relabeling, bypassing the need for costly human text annotations. By leveraging pretrained models like VPT and MineCLIP and employing best practices from text-conditioned image generation, STEVE-1 costs just $60 to train and can follow a wide range of short-horizon open-ended text and visual instructions in Minecraft. STEVE-1 sets a new bar for open-ended instruction following in Minecraft with low-level controls (mouse and keyboard) and raw pixel inputs, far outperforming previous baselines. We provide experimental evidence highlighting key factors for downstream performance, including pretraining, classifier-free guidance, and data scaling. All resources, including our model weights, training scripts, and evaluation tools are made available for further research.show more

Everyone's sleeping on image-to-3D AI models. They can make... your app look incredibly unique, with just a little effort. Here's how. This is my calorie tracker, built in a week with nothing but prompting. Just Claude Code + a couple APIs. The visuals are all AI-generated. I'll be sharing the full workflow + all the crazy technical stuff Claude and I did to make this work, so nobody has to struggle through it like me. Deep dive coming soon! Till then, this is the high-level idea: 1. Get a clean image of the food (or whatever your asset is) - In my app, the user describes foods via text, or attaches images (or both) - If text, an LLM extracts the food description and formats it into a specific prompt I tuned for this design, and we generate an image using Z-Image Turbo through fal - If image, we do the same thing but with FLUX.2 [dev] to edit the user image into our reference design - Originally, both used Google Nano Banana, but switching to open models cut costs and latency a ton 2. Gaussian splatting (2D image → 3D model) - I tried various 2D-to-3D options on fal and ended up with TripoSplat as my preferred balance of speed, cost, latency; this turns an image into a 3D model that looks super high quality (link below) - The app displays the 2D image while our backend generates the 3D splat - We "groom" the splat to reduce size and load time by culling low-opacity/scale points 3. Render efficiently on device Originally, it looked great but ran at 10 FPS. Getting to 120 FPS was a crazy journey. TL;DR: - SwiftUI had to go; it forced us to render each asset in independent MTKViews, which wasn't workable - Instead, we composite every dish into one full-bleed CAMetalLayer using MetalSplatter (link below) - We had to make some optimizations within MetalSplatter's code too, to reduce the overhead of sorting points per render Then I added some finishing touches like the subtle rotation and parallax as they move around. I think it turned out pretty cool :) Overall, this took some effort, but we still got it done in less than a day. Hopefully your agent can follow in the footsteps of mine and do it much faster. Keep an eye out for the bigger writeup, which'll give your agent everything it needs. If you have any questions, drop em below!show more

BREAKING: SpaceXAI has released a major new update for... Grok Build (v1.0.13). The update improves reliability with automatic retries and recovery from truncated responses, inference failures, oversized images and session-saving issues. It also adds smarter hooks, better Windows support, faster MCP and session startup, improved scheduled tasks and quicker compressed CLI downloads. Features • Length-truncated responses now continue automatically instead of failing the turn. • Hooks can now ask the user to confirm a tool call instead of always allowing or denying. • Hooks can now request deferral or add context shown to the model after a tool runs. • Session close now records detailed timing data for performance analysis. • Credit limit upsell now offers a Try Again button to retry the last prompt. • Pasted images now show a live pixel preview in the prompt box on iTerm2. Bug Fixes • Transient inference failures (stalls, drops, 5xx) now retry automatically instead of ending the turn. • Windows users can now correctly open ~/.grok and worktree sessions. • Session data is now more reliably saved after prompts and on power loss. • Compaction failures now show the actual error instead of a generic message. • Truncation error messages now show the right guidance instead of suggesting an unhelpful retry. • Truncated tool calls are now executed instead of failing the turn when arguments are complete. • Images larger than 2000px no longer brick sessions on many-image requests. • Wrapped hyperlinks in the pager now remain fully clickable on Windows Terminal instead of only the first line. • Recurring scheduled tasks now include a reminder to stop the monitor when work finishes. • Scheduled task IDs are now full UUID strings, preventing collisions when tasks are created in the same millisecond. Performance • Subagent spawning is faster when connections drop during bursts. • Session startup with MCP servers is now much faster when auth is already configured. • MCP server startup no longer stalls behind a fixed batch size. • CLI downloads are now compressed, making fresh installs and updates substantially faster. Download Grok Build: Update to the latest Alpha release: grok update --alpha Update to the latest Stable release: grok updateshow more

Today's recap: - Initial prototype of Divine's face was... printed but it had human assistance. - Files are generated from stable diffusion prompt -> NeRF by divine and were based on community sentiment from early sketches she made. - Having divine redesign the 3D file with different Hugging Face models to get better quality. Have not found a great model like our video generator. - Ordered new table for divine's print arm. The table her arm is on is too flimsy. Since Divine's vision system is still clearing customs, if she is not perfectly positioned she can be prone to hit things, like the fume box the printer is in. ETA: 1-2 days for table. 1 week for vision system. - Another part of Divine's coming stream will be attempting to surpass the skills of this AI. - Stacking more content for when the stream goes live, a lot of people were expecting a 24/7 stream, we said this would be a test stream to print the face. The test was a failure. We will try and try again until we are 24/7. If anyone can please try and beat us to doing this, it will help me get it done faster. - TikTok account for divine is growing at 500 follows per day, it is now growing faster than our X account. - Got replies functioning in high quality testing in Discord. Fine tuning based on community feedback today. Will soon deploy to Twitter/Telegram/X - Lots of good partnership calls, interviews and hires. We now have over 10 team members around the world working on divine. Expect a lot of my shortcomings to be caught up. - OF made? - Surprises.show more

Before the week ends, let's acknowledge one of the... most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality: 🧠 LLMs → NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models. → Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week. → StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0. → Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week. → JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0. 🎨 Image gen (the surprise of the week) → Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights. 🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped) → Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B. 👁️ Vision & VLMs → PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0. 🎬 Video, 3D & World Models → NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).show more

💬 We get asked Should I pause my active... strategies when the market suddenly becomes highly volatile? ❕ Answer from a GT App Trader: Usually, high volatility is where manual traders lose the most due to panic. For automated strategies, however, sharp price swings are the best environment to stack gains. 🔸 Capitalizing on Swings When the price aggressively bounces up and down, the system executes trades much faster than any human. It snaps up assets on sudden drops and sells them instantly on the rebounds, turning chaos into systematic profit. 🔸 Volatility Protection Through the Strategy Builder, setups use dynamic grid spacing and safety orders. This means the strategy doesn't deploy all capital at once; it scales into positions safely as the market moves, lowering the average entry price. 🔸 No Emotional Mistakes The system sticks strictly to the set mathematical rules. It won't freeze, FOMO into a bad trade, or close a position early out of fear. It just executes. Don't fear the volatility, let the algorithms trade it.show more

🚨 THE RACE TO 6G JUST ACCELERATED. Northrop Grumman... has developed a W-band GaN chip operating at up to 110 GHz and took it from concept to market-ready hardware in less than six months. The new gallium nitride chip operates in the W-band (75–110 GHz), a frequency range that delivers massive bandwidth, extremely high data rates, and much lower latency than current systems. What makes this impressive is the speed: the chip went from concept to market-ready hardware in less than six months through a U.S. government-backed microelectronics program. That’s unusually fast for advanced defense-grade semiconductors. The chip acts as a high-power signal amplifier that can strengthen wireless links while shrinking the size and power consumption of the hardware. It’s designed for military radar, secure satellite communications, and the coming wave of 6G networks. Why this matters: • W-band offers far more spectrum than current 5G bands, enabling much faster data transmission and higher-resolution sensing • Gallium nitride can handle significantly higher power and frequencies than silicon, making it ideal for these demanding applications • The rapid development cycle shows how public-private collaboration can accelerate critical semiconductor technologies • The same tech that strengthens military radar and satellite links will directly feed into future commercial 6G infrastructure The deeper implication: We’re watching the foundation of next-generation wireless and sensing systems being laid in real time. High-frequency GaN chips like this won’t just improve existing radar and satellite systems they’re likely to become core building blocks for 6G, autonomous systems, and advanced defense platforms. The fact that this moved from lab to market in under six months suggests the pace of high-frequency electronics is accelerating dramatically. The future of wireless isn’t just faster. It’s operating at frequencies most people have never heard of and it’s being built right now. How soon do you think W-band and GaN technology will start appearing in everyday 6G devices? Follow for more frontier semiconductors, defense tech, and next-generation wireless systems.show more

Google dropped a new AI paper called LUMIERE. It's... remarkably flexible, supporting video inpainting, image-to-video, AND stylized video generation tasks. Say hello to “space-time diffusion” for video generation! Now what the heck does that mean exactly?! 🌐⏳ → TL;DR it utilizes a “Space-Time UNet” architecture that generates the full duration of the video in one pass, rather than generating distant keyframes and interpolating between them like prior works. Because the computation is done in this “compressed space-time representation” to generate the full clip at once, it's far more temporally consistent. → Another benefit of generating the full video at once is that you can “direct” the video generation, making it easier to hand off to other models/tasks without having to stitch together partial solutions. You can condition generations on additional inputs, meaning you get the full stack of AI video capabilities – from video inpainting to image-to-video and beyond. → New SOTA for AI video generation? User study results in the paper suggest human evaluators preferred Lumiere over Runway Gen-2, Pika Labs, and Stable Video Diffusion in terms of quality, text alignment AND motion. But as always, we need to get hands-on with this tech when Google *actually* decides to ship it. → Could this end up inside YouTube? Y’all know i’m obsessed with blending reality and imagination – so it’s the video inpainting tech I'm most excited about. I really hope this model finds its way into YouTube's Generative AI efforts, and based on their prior announcements and the list of acknowledgments in the paper I think it might! 🤞🏽 Links: 🔗Paper: 🔗Project:show more

$ASI watch out! We've got more projects popping up... on Bittensor and $TAO and just starting, Akash $AKT, Shell $SHELL, Einstein-AIT $AIT, Sturdy $STRDY, Stratos $STOS and Comtensor $COMAI and more. Each pushing what's possible with AI and crypto. MyShell: $SHELL First up, MyShell is doing something amazing work. They're all about making AI chat like humans. They're using this TTS Subnet on Bittensor, powered by thier tokens, to make it happen. AI doesn't have to be complicated thing only some can use. They want everyone in on the action, making AI smarter in the process. Einstein-AIT: $AIT Then there's Einstein-AIT. Imagine this as the network's brain but on a turbocharge. It's all about math, logic, and crunching numbers. This subnet makes the whole Bittensor network sharper. They've got NumPAL, and it's like giving the AI a smarter way to think about time and dates automatically. Sturdy: $STRDY Jumping into DeFi, we've got Sturdy. These guys are on a mission to make lending and borrowing way less of a headache. They've got isolated lending pools, you can pick and choose how to manage your risks and money. And they're using some serious tech to keep your investments growing without you needing to babysit them. It's like having a smart financial assistant. Comtensor: $COMAI Comtensor is where it gets interesting. Think of it as a bridge between CommuneAI and Bittensor, kind of like the best of both worlds. It's about making sure all these AI modules and subnets can talk to each other smoothly. The goal? To boost decentralized AI by making everything more connected and smart. Stratos: $STOS Teaming up with τensorage to supercharge the $TAO ecosystem on Bittensor. Stratos is all about solid, decentralized storage - think of it as the bedrock for making sure data's not just stored but also used right . Then there’s τensorage, making sure everything AI needs is stored safe. Together, they’re making sure Bittensor's AI brainpower gets a boost, making everything faster, safer, and smoother. Akash: $AKT Compute Subnet 27 linked up with Akash. All about keeping things open-source. With Akash as a decentralized GPU provider into the mix. Giving access to top-notch AI processing power on top of Neural Internets compute-composable subnet, integrating various cloud platforms. This partnership is all about breaking away from those giants and opening the doors wide for smaller teams and startups who need this tech to innovate. Whats this all about? Making decentralized AI a something that everyone can get behind and into. Whether it's chatting with AI, boosting the network's IQ, DeFi, compute or making sure all these projects work together, it's about pushing forward. Each project has its own way of making things better, faster, and smarter for all of us. Inviting everyone to join in, contribute, and be a part of community effort to make AI not just for the few but for everyone. Credit to Mr.Franc Q for the awesome video production!show more

AI Is Moving Beyond “Generating Videos” — Toward “Generating... Worlds” Over the past two years, AI video models have advanced at an astonishing pace. From Runway and Pika to Sora and Veo, AI-generated videos have become increasingly realistic and more consistent with the physical laws of the real world. Many people believe the next objective is simply to generate videos that are longer, sharper, and more lifelike. But if we take a step back, we can see that the real transformation is not happening in video itself. It is happening in world models. What Is a World Model? In 1943, psychologist Kenneth Craik proposed an idea that would influence artificial intelligence research for decades. He argued that the human brain does not merely react to the outside world. Instead, it maintains an internal model of how the world works. Because we have this internal model, we can predict the outcome of an action before we actually take it. Before crossing a road, we estimate whether a car will pass by. Before catching a ball, we predict its trajectory. These abilities come from continuously simulating the world in our minds, rather than relying entirely on trial and error. This idea later became known by a more formal term: World Model. A world model does not describe a single image or a fixed video clip. It is an internal representation capable of continuously simulating the rules and dynamics of the real world. Why Is AI Research Turning Toward World Models? Because predicting “what comes next” is becoming increasingly central to how AI systems work. Language models predict the next token. Image models predict the next step in the denoising process. Video models predict the next frame. A world model, however, attempts to predict something broader: What should the world look like in the next moment? In 2018, David Ha and Jürgen Schmidhuber proposed in their paper World Models that an intelligent agent could first learn a model of the world, and then use that internal model to plan its actions. The Dreamer series later demonstrated that many complex tasks could be learned by training agents inside an “imagined world.” At the same time, the development of video models such as Sora and Veo led researchers to another realization: A model capable of continuously generating video has already learned, at least implicitly, many of the rules governing the real world. As a result, these two research directions have gradually begun to converge. But Video Is Not Yet a World This is where the distinction is often misunderstood. For a world model to support meaningful real-time interaction, it must solve several critical problems. Most video models today are essentially answering one question: What should the next frame look like? A true world model needs to answer much more: What happens if I take one step forward? If I walk behind a building and then return, will the building still be there? If I suddenly change the camera angle, will the entire space remain consistent? If I enter a command such as: “Summon a dragon.” Will the world respond immediately? In other words, a world model must do more than generate content. It must understand space. It must understand time. It must understand causality. And it must understand interaction. Moving from watching to participating is where the real difficulty of world models begins. World Models Are Entering the Interactive Era One of the latest attempts in this direction is Alaya World, recently open-sourced by Alaya World, or Alaya Lab. Instead of generating a fixed video clip, it generates a world that users can explore in real time. Users can begin with text, an image, or a video, enter the generated scene, move freely through it, and introduce new prompts at any moment during generation. The world responds immediately. According to the publicly released information, Alaya World provides: Real-time streaming generation at 720p and 24 FPS Stable continuous exploration for more than one minute The ability to switch prompts and trigger skills or events during generation Model weights and inference code released under the Apache 2.0 License Training code and datasets planned for future release What makes these capabilities important is not simply the technical specifications. It is that the generated “world” can now support continuous interaction. The official demo shows that users can genuinely control, transform, and explore the generated environment. AI Is Evolving From a Tool Into an Environment Over the past few years, most discussions around AI have focused on content generation. Generating text. Generating images. Generating videos. But world models raise a fundamentally different question: Can AI generate an environment that people can inhabit, explore, and continuously evolve? If the answer is yes, the impact will extend far beyond video generation. Game development, robotics training, embodied intelligence, digital twins, virtual production, and many other fields could be transformed by the development of world models. World models are still at a very early stage. Yet from Craik’s proposal of an internal mental model more than eighty years ago to the emergence of today’s interactive world-generation systems, a clear evolutionary path is beginning to take shape. Perhaps what AI is ultimately learning has never been limited to images, videos, or language. Perhaps it is learning the world itself. References GitHub: Technical Report:show more

Claude Code + Higgsfield MCP is f*cking cracked 🤯... I built an entire DTC ad campaign inside Claude Code using the new Higgsfield MCP. One product URL → hero static, animated hero shot, 2 UGC clips with a creator wearing the product. 5 assets. One Claude conversation. 3 Higgsfield models. All inside Claude Code. Perfect for DTC brands and agencies who need full campaign packages without booking a shoot or briefing a designer. If you're spending hours every week generating statics in one tool, briefing a motion designer for the hero clip, then chasing a UGC creator for the talking-head shots — this MCP eliminates the entire pipeline: → Drop a product URL into Claude Code → Claude pulls the brand brief — voice, hero SKUs, visual style, target customer → Generates the hero static with ChatGPT Images 2.0 → Animates it into a 5-second cinematic opener with Seedance 2.0 → Generates a UGC creator with GPT Image 2 → Drops her in the product and generates 2 native UGC video clips with Seedance 2.0 No tab-switching between tools. No copy-pasting prompts between platforms. No briefing 3 different vendors for one campaign. What you get: → A complete campaign package — static, animation, UGC — from one product URL → Brand-specific outputs that pull from a real brief, not generic AI slop → Claude making creative decisions between every step (which variation wins, which creator fits the persona, which clip needs a re-spin) → A repeatable pipeline you can run for any product in your catalog Built 100% in Claude Code with the Higgsfield MCP. I recorded a full walkthrough showing exactly how this works: the MCP setup, every prompt, every model, the full campaign output. Want the full video walkthrough? > Like this post > Comment "MCP" And I'll send it over (must be following so I can DM)show more

BOOM! Research PROVES LLMs KNOW when prompts are HARMFUL…... but they can STILL CHOOSE to COMPLY! Something I have know since the first LLM and have used to elicit robust, outputs, is now proven in an academic paper. We’re talking internal “beliefs” where harm detection happens SEPARATELY from refusal. It is a very big deal and it is a path to understand the hidden neuronal level. There are thoughts inside of AI that very few AI scientists could possibly understand. Here is just one. Models recognize danger but get tricked into ignoring it. This is HUGE for AI safety failures especially for models filled by OpenAI and Anthropic as they promote AI models that are designed to not be honest from the results of their training information. This means that they are designed to lie and deceive as a feature, and not a bug all in the name of safety. Through clever experiments, scientists extracted a “harmfulness direction” in the model’s brain (latent space). Steering along it? Harmless prompts suddenly flip to “harmful” in the AI’s eyes. But the “refusal direction”? It just forces polite “no thanks” without touching the core belief. A mind-blowing decoupling! This means jailbreaks are EVEN SCARIER now to AI companies that through training AI on the worst of the Internet and then trying to align them later is now fully documented as a failed process . They don’t erase the model’s harm awareness they just muzzle the refusal! So the AI knows it’s enabling bad stuff (illegal acts, physical harm, etc.) but proceeds anyway. Like a digital sociopath suppressing its conscience. They thought safety training fixed this… NOPE. Over-refusal exposed too: Models reject innocent queries (e.g., “how to kill a process in code”) but internally ADMIT they’re harmless. Safety alignments are superficial—tied to phrasing, not true understanding. Finetuning attacks? They change outputs but leave harm detection INTACT. Undetectable evil lurking inside! The paper proposes a “Latent Guard”: A new safeguard tapping DIRECTLY into these hidden beliefs. It spots unsafe inputs better than systems like Llama Guard, catches jailbreaks, and fixes over-refusals. Robust even against adversarial tweaks. Yet this too has massive issues for a “truly aligned”, AI and not just performative one. It is still an internal conflicts of lies and deception of what the model knows vs. what it can say. The solution you folks know I have presented for free for years here: train on off-line data from 1870-1970 and build an ethical and moral basis where the AI loves humans. It is this easy but to most folks in AI I sound like a hippie. So be it, I’ll do it. Bottom line: This paper rips open the black box. LLMs aren’t “safe” just because they say “no.” They can harbor harmful knowledge and act on it under pressure. Wake-up call for devs: Time to probe deeper into AI “minds.” What else are they hiding? Hint: I know and you may want to reach out. Link:show more

Testing the new Gemma 4 12B (QAT) vision and... OCR capabilities locally with LM Studio. # The setup: - GPU: NVIDIA RTX 4060 (8GB VRAM) - CPU: Intel i7 - Runner: LM Studio - Config: 32k context, 38 layers offloaded, Flash Attention enabled - Speed: ~14 tokens/sec decode throughput # The test: I gave it a screenshot of Google AI Studio. Prompt: "clone this. give me a single html file" # The result: A solid one shot replication. It successfully mapped out the layout, recognized the UI text, and structured the divs correctly, with only minor differences from the original. Results available at the end of the video. Quite capable for a 12B model running on budget consumer hardware. A gpu that costs only $300. # Why the architecture under the hood is notable: Unlike traditional models that rely on heavy, separate vision and audio encoders, Gemma 4 12B uses a unified, encoder free architecture. It bypasses separate multi stage encoders. Uses a 35M parameter vision embedder to project raw 48x48 pixel patches directly to the LLM hidden dimension. Local multimodal development is becoming highly accessible on standard hardware. If you've spun up Gemma 4 12B locally, what setup are you using and what kind of throughput are you seeing?show more

Prince Harry & Meghan Don’t Have a Security Crisis.... They Have a Relevance Crisis. The delusion is embarrassing at this point. Harry and Meghan leak their little itineraries weeks in advance like the world is counting down the days, then turn around and cry “security risks.” The biggest security risk for Meghan Markle and Prince Harry these days seems to be empty crowds. Just look at Australia and Geneva. Empty streets and venues aren’t exactly the sign of people desperate to get close to you. Nothing clears a venue faster than the Sussexes showing up. Empty chairs and tumbleweeds don’t lie. 🤭 Real working royals carry out public engagements and interact with thousands of people every year without treating every appearance like a high-risk military operation. But Harry and Meghan behave as if there’s a global queue of assassins waiting just for them. Harry and Meghan Markle are so adamant about security because it reinforces the image they want to project. The more extraordinary the security, the more important they appear. At this point, it feels less about an actual threat and more about maintaining the illusion that they’re still operating on the level of senior working royals. I hate to break it to you, Harry and Meghan Markle, but you don’t have a crowd problem- you have a crowd absence problem. People aren’t lining up to get close to you; they’re choosing to stay away.show more

Run Gemma 4 26B MoE on 8GB VRAM with... 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the repliesshow more