xAI Grok 4.5 Model Capabilities

There's now a free IC-LoRA that lets you reposition a camera in an AI video by literally drawing a sphere and picking a new angle—instead of praying your text prompt gets interpreted the way you meant it. Load it into ComfyUI, grab the companion node, and you're testing viewpoint changes on your own footage within the hour. Precision instead of vibes.

Continuing my technical assessment prep — today's focus was the sliding window pattern. Worked through 6 problems: Minimum Size Subarray Sum, Minimum Window Substring, Sliding Window Maximum, Permutation in String, Find All Anagrams in a String, and Subarray Product Less Than K. Sliding window is one of those patterns that's genuinely worth internalizing deeply — expand and shrink a window over the array while tracking a running condition, avoiding the need to re-scan from scratch. It shows up constantly once you can spot it. Day 48 of my 100 Days of AI series. #100DaysOfCode #LeetCode #DataStructuresAndAlgorithms #SoftwareEngineering #TechnicalInterview #100DaysOfAI #BuildInPublic

Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks. Most AI labs move at glacial speed. xAI is moving at Starship cadence. Insane pace. While others do safety theater and endless alignment meetings, you're just shipping god-tier models every couple weeks. The real acceleration begins now. This is how you win.

My dog is faster than my object detection model's confidence threshold. This project was built to fix that. Tuffy has one job in this house: look adorable. He has recently added a second job: attempting a full sprint escape every time the front door opens for more than two seconds. So I did what any reasonable engineer would do and built a real-time detection system to catch him in the act. DogVision AI is a YOLO26n model fine-tuned on a custom dog dataset, running through a live webcam pipeline with tracking, a grayscale feed, an edge map, and a heatmap, all composited into one dashboard. Training stats that took less effort than photographing Tuffy standing still: Precision 98.5 percent Recall 97.5 percent mAP50 99.4 percent mAP50-95 72.7 percent Weights under 6 MB Getting the model to converge took about 99 epochs. Getting Tuffy to sit in front of the camera for longer than three seconds is still an unsolved research problem. Code, training curves, and Tuffy's escape attempts are in the repo, link in comments. #YOLO #ObjectDetection #ComputerVision #MachineLearning #AI #DeepLearning #OpenCV #PyTorch #Ultralytics #SideProject #DogsOfLinkedIn

Hy-Embodied-RxBrain-1.0 is here: a ~6.2B embodied cognition model that can reason about a scene, imagine what happens next, and plan robot subgoals. 🚀 License: Apache 2.0. 🤖 https://lnkd.in/eTY-yR3R 🧠 Scene understanding: answers questions over images and multi-frame video. 🔮 World modeling: generates near-future frames after an action. 🦾 Subgoal planning: breaks a task into steps, pairing each action with the visual goal state it should reach. ⚙️ Unified MoT: text, vision, and generation use modality-specific paths inside one autoregressive model, instead of separate towers. #EmbodiedAI #WorldModel #RobotLearning #MultimodalAI #SpatialReasoning #ModelScope

Grok 4.5 is keeping the model race interesting weeks after launch. Beyond its coding and agentic performance, SpaceXAI’s emphasis on efficiency is the part worth watching. In agentic systems, how efficiently a model navigates tasks, tools, and decisions without burning excessive compute is increasingly important. As AI moves into practical deployment, that efficiency is becoming a clear product advantage.

How I used Claude AI to build an interface able to perform manual design for a ribbed slab In this video, you get to learn why decisions are made when designing a ribbed slab

🏆 Poolside ships Laguna S 2.1: 118B-param open-weight MoE that beats DeepSeek V4-Pro Max on SWE-bench (59.4% vs 55.4%). Runs on a single DGX Spark. "The West's most capable open-weight coding model." https://lnkd.in/dVD3k2vX 📧 Enjoying this? Subscribe to my daily AI newsletter — email julio.rodriguezmartino@gmail.com with subject: SUBSCRIBE TO AI NEWS

50 references steering one Seedance 2.5 generation, inside Runway. The reference ceiling in AI video is about to move. Seedance 2.5 is on its way to the Runway platform. The headline numbers: → Up to 50 references shaping a single output → Clips reaching 30 seconds → 7 days of uncapped access at launch for new Max plan subscribers The reference count is the technical story. Identity, product, and style control in generative video has always been bounded by how much steering material one generation can hold. A 50-input ceiling brings model direction closer to how a production brief is actually written: multiple boards, multiple anchors, one deliverable. Longer clips widen the edit margin. The reference stack decides whether the output matches the brief.