Crazy ML (@crazyMLguy) header
Crazy ML (@crazyMLguy) profile picture

Machine Learning & Backend ๐Ÿ˜Ž Founder of Humanity backed by Investors from Google, SF and Hong Kong.

Mountains
Joined June 2023
Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)Photo by Crazy ML (@crazyMLguy)
I love Cycling, Machine Learning and Backend.
8
784
๐—Ÿ๐—ฒ๐—ถ ๐—๐˜‚๐—ป ๐—ฎ๐˜ ๐—”๐—ฃ๐—˜๐—–: ๐—”๐—œโ€™๐˜€ ๐—ฅ๐—ฒ๐—ฎ๐—น ๐—•๐—ฟ๐—ฒ๐—ฎ๐—ธ๐˜๐—ต๐—ฟ๐—ผ๐˜‚๐—ด๐—ต ๐—ช๐—ถ๐—น๐—น ๐—•๐—ฒ ๐—ถ๐—ป ๐˜๐—ต๐—ฒ ๐—ฃ๐—ต๐˜†๐˜€๐—ถ๐—ฐ๐—ฎ๐—น ๐—ช๐—ผ๐—ฟ๐—น๐—ฑ Not in chat. In phones, cars, and factories. Xiaomiโ€™s factories already run fully automated with AI. Next step: industrial agents that can produce, spot problems, and improve on their own. Their embodied robot just did an 8-hour factory shift and charged itself. AI moving into the physical world is the real game-changer. #AI #Xiaomi #EmbodiedAI #PhysicalAI #APEC
1
23
๐—ง๐—ต๐—ฒ ๐—ช๐—ผ๐—ฟ๐—น๐—ฑโ€™๐˜€ ๐—™๐—ฎ๐˜€๐˜๐—ฒ๐˜€๐˜ ๐—ฆ๐˜‚๐—ฝ๐—ฒ๐—ฟ๐—ฐ๐—ผ๐—บ๐—ฝ๐˜‚๐˜๐—ฒ๐—ฟ ๐—ฅ๐˜‚๐—ป๐˜€ ๐—ผ๐—ป ๐—–๐—ฃ๐—จ๐˜€ ๐—ก๐—ผ๐˜ ๐—š๐—ฃ๐—จ๐˜€. Chinaโ€™s LineShine is doing something unusual in the age of GPU-dominated AI. It has secured the No. 1 spot in the TOP500 with 2.198 exaflops, powered by 13.78 million CPU cores. But the more interesting achievement is what researchers are doing with it: โ€ข Simulating a silicon crystal containing 200 million atoms. โ€ข Running large-scale quantum-level material simulations. โ€ข Achieving 96.6% weak-scaling efficiency in a 100-million-atom study. This could help scientists study materials and battery technologies at a scale that was previously out of reach. The bigger story: the future of computing isnโ€™t just about faster AI chips. Itโ€™s also about solving scientific problems at an entirely different scale. #Supercomputing #AI #HPC #Research
1
4
31
SGLang Low-Latency Mode: Making LLM Inference Faster SGLang, NVIDIA and AMD are working on a new approach to reduce LLM inference latency. Instead of relying only on standard kernels, they introduce specialized GPU kernels that reduce launch overhead, overlap computation with communication, and optimize small-batch decoding. Key results: Up to 1.38ร— faster token generation on NVIDIA GB300. Up to 1.56ร— faster on AMD MI355X. NVIDIA optimizations improve per-user decode speed by 12โ€“38% in the tested workloads. The interesting part is the plugin architecture: specialized kernels can integrate with SGLang without putting every hardware-specific optimization into its core codebase. Code: NVIDIA: github.com/sgl-project/sgl-nโ€ฆ AMD: github.com/sgl-project/sgl-aโ€ฆ Read: lmsys.org/blog/2026-10-18-sgโ€ฆ #LLM #Inference #SGLang #GPU #AI
3
3
16
787
Hackers are now using LLMs to continuously scan software updates, spot newly fixed bugs, and build attacks before teams can apply the fix โ€” all without human involvement. Thatโ€™s according to the latest report from the Google Threat Intelligence Group. We're helping defenders counter automated threats with Google AI Threat Defense, which combines Gemini and other frontier models to find and fix flaws before attackers can exploit them.
377
2,795
658
17,623
1,125,375
BridgeKV: Faster LLM Inference Without Losing Large-Model Quality Thatโ€™s the idea behind BridgeKV. It translates the large modelโ€™s KV cache into a format the smaller model can use, allowing the smaller model to handle decoding. Key results: Up to 3ร— faster requests in a real A40 GPU test. Retains much of the larger modelโ€™s quality. Works across six model pairs and five datasets. A promising approach to making LLM serving faster and more efficient. Paper: arxiv.org/abs/2611.11761 #LLM #Inference #AI #MachineLearning
8
57
1,795
Crazy ML retweeted
Today we're open-sourcing HubEgo-100K: the most diverse egocentric dataset on ๐Ÿค— @huggingface. 100,000 hours of skilled tasks from 100,000+ different Hub contributors. Free for research and commercial use ๐Ÿงต
111
101
39
460
340,717
Wow cool and can we able to call unrestricted and uncenscored models too like these ones because if so then it would be game changer.
We're teaming up to accelerate open-source Al. OpenRouter is now integrated into Hugging Face directly. Same token, same endpoint, plus live model stats on the Hub and an inference widget we redesigned together. The best of open, in one place. huggingface.co/blog/openroutโ€ฆ
1
83
LingBot-Touch: Giving Robots a Sense of Touch. Robbyant has introduced LingBot-Touch, a tactile foundation model designed to help robots understand physical contact across different sensor types. Key highlights: Trained on nearly 5,000 hours of tactile data and 1.33 billion frames. Unifies magnetic, piezoresistive and optical tactile sensors. Improves robot manipulation success by up to 26 percentage points across 26 real-world tasks. Transfers to previously unseen tactile sensors without retraining. The bigger idea: Robots need more than vision. Touch could be the missing piece for truly dexterous manipulation. Website: technology.robbyant.com/lingโ€ฆ GitHub: github.com/robbyant/lingbot-โ€ฆ Models: huggingface.co/collections/rโ€ฆ Research: arxiv.org/abs/2611.12806 #Robotics #EmbodiedAI #AI #FoundationModels #TactileSensing
2
1
7
318
Wow so cool, kudos @UnslothAI
Unsloth now has computer use for any local model! It's still beta but we had Qwen3.8-27B make an expense report from receipts & used GIMP to scale photos, add captions & export it! Also we added driving Unreal Engine, Blender & more to local models! github.com/unslothai/unslothโ€ฆ
1
24
EVAThink: A Video AI Model That Knows When to Think and When to Answer Researchers from East China Normal University and Huawei introduced EVAThink, a training method that helps video AI models decide when deeper reasoning is actually needed. The idea is simple: Enough visual evidence? Answer directly without wasting tokens. Limited visual evidence? Think through the available clues before answering. How does it learn? Researchers create video pairs with the same question but different levels of visual evidence, then use two-stage reinforcement learning (GRPO) to teach the model when to use each response mode. Results: Trained using just 6.7K samples. Achieved 65.6% average accuracy across six video benchmarks with 64 frames. Outperformed both direct-only and think-only modes. Reduced output length by 62% compared with always using reasoning. Built on Qwen3-VL-8B-Instruct. The bigger idea: Better video AI isn't about thinking more. It's about knowing when thinking is worth the extra computation. Paper: arxiv.org/abs/2611.13907
1
12
420
Evolving Infra: An AI Inference System That Automatically Adapts to Agent Workloads MIT researchers have introduced Evolving Infra, an LLM serving framework designed for dynamic AI agents like Claude Code and computer-use agents. Instead of using fixed GPU allocation, it automatically adjusts infrastructure based on workload demands. Key innovations: Dynamic GPU allocation: Moves GPUs between prompt processing (prefill) and token generation (decode). Smart load prediction: Predicts workload demands to distribute requests efficiently. Intelligent KV cache management: Keeps useful conversation history and removes cache data that agents no longer need. Reported results compared with NVIDIA Dynamo: Up to 4.1ร— lower end-to-end latency. Up to 14ร— lower P95 time to first token. Up to 2.7ร— higher throughput. The idea: Build inference infrastructure that adapts to how AI agents actually work instead of relying on fixed configurations. GitHub: github.com/mit-han-lab/Evolvโ€ฆ Paper: arxiv.org/abs/2611.12304 #AIInfrastructure #LLM #Inference #AI
2
4
27
892
MODHARNESS-RL: A New Way to Make AI Agents Improve While They Learn Researchers from Penn State, LinkedIn, and UC Santa Barbara introduced MODHARNESS-RL, a framework that improves both an AI model and the system that guides it. It works through three mechanisms: Model training: Updates the model's weights using GRPO reinforcement learning, based on task performance. Compression: An LLM creates executable code rules that decide which parts of previous conversations and tool results are important enough to keep in context. Steering: The LLM creates more executable rules to detect recurring mistakes and provide corrective instructions. These rules are tested against previous interactions before being accepted. Steering rules can evolve as the model learns, while the compression rules remain fixed during training. Results: Up to 24% improvement over fixed-harness training on CodeSearch, with additional experiments on ALFWorld and BIRD SQL. The important distinction: The model learns through weight updates, while the executable rules improve how it uses context and handles mistakes. Paper: arxiv.org/abs/2611.18533
1
1
3
331
Open Vision Banana: Turning Image Generation Models Into AI Vision Systems Researchers have introduced Open Vision Banana (OVB), an open project that turns image generators into models that can understand and analyze images. Instead of just generating pictures, the same model can: Identify and segment objects in images. Understand which object a text description refers to. Estimate depth and surface geometry. Generate and edit images. What's impressive? Built on FLUX.2 Klein 4B and 9B. Trained using 2.1Mโ€“3.8M public data pairs. Achieved 73.9 mIoU on Cityscapes zero-shot, compared with Vision Banana's reported 69.9. A new technique called Legend Ensembling improves segmentation accuracy without retraining. The researchers have released the code, training pipeline, datasets and model checkpoints. The bigger idea: One image-generation model can become a multi-purpose computer vision system. GitHub: github.com/OpenVisionBanana/โ€ฆ Project: openvisionbanana.github.io Paper: arxiv.org/abs/2611.15844
1
3
17
564
Tencent's FlashMotion Generates 3D Human Motion in Just 400ms Tencent has introduced FlashMotion, a lightweight AI framework that converts text, music, and speech into smooth 3D human movements. What's impressive: Generates 200 frames (6.7 seconds) in just 400ms on an iPhone 17. Supports motion inpainting, filling gaps between existing movements. Achieves 173,172 FPS on an RTX 3090 with one sampling step. Up to 13.6ร— faster than previous methods in reported benchmarks. Fast, lightweight AI motion generation, even on smartphones. #Tencent #AI #3DAnimation
1
29
How Qwen4-Max Trained a 2.4T-Parameter Model Without a Single Loss Spike Qwen4-Max uses three key architectural improvements to make large-scale training more stable and efficient: GDN + QSA: Combines efficient memory of past context with smart retrieval of relevant information. Gated Residuals: Uses four parallel information pathways to improve information flow and training stability. N-gram Embeddings: Adds extra context from nearby tokens to improve the model's understanding without significant extra compute. The bigger idea: Better architecture can make massive AI models more efficient and stable to train. #Qwen #AI #LLM
1
38
Qwen4-Max Introduces Live Delivery: AI Agents That Build, Deliver, and Improve Work Step by Step Usually, you give an AI agent a big task and wait until everything is finished. Qwen4-Max takes a different approach. With Live Delivery, agents can: Deliver a working first version early. Keep improving it and releasing updated versions. Take user feedback and adjust their plans. Let users preview, download, and review each version. Example: Ask an AI agent to build a website. It delivers a basic working version first, then adds features and improves the design based on your feedback. Alibaba has also open-sourced the plugin, currently supporting Codex CLI. The idea: Don't wait for the final result. Build, deliver, and improve continuously. GitHub: github.com/QwenLM/Qwen-Live-โ€ฆ #Qwen #AI #AgenticAI #OpenSource
2
4
425
Tencent Launches Hy4 Flash: 224B Parameters, Just 12B Active Key features: 224B total parameters, 12B active. Multimodal support. 1M-token context window. API pricing per 1M tokens: Cached input: $0.003 Input: $0.075 Output: $0.30 That's up to 15ร— cheaper than Hy4 Available on WorkBuddy, Yuanbao, and WeCom, with OpenRouter integration coming soon. #Tencent #AI #LLM
27
Huawei Open-Sources CloudRobo-WM1.2: A New World Model for AI-Powered Robots Huawei Cloud has released CloudRobo-WM1.2, an embodied AI world model built for its Ascend AI hardware. The idea is simple: help robots understand their surroundings, predict what will happen after an action, and decide what to do next. Key innovations: PPA: Helps the model build a better understanding of the environment using multiple types of input. DualAct: Combines world prediction with robot action generation. Adaptive training: Learns from data at different levels of detail. Cross-robot learning: Helps transfer motion knowledge across different robot designs. Four inference modes: Flexible ways to combine world prediction and action generation. Huawei reports that it ranks No. 1 on RoboTwin 2.0 and LiberoPro 2.0 benchmarks. Huawei is also opening its training framework and Ascend optimization practices to developers. The bigger idea: Robots that don't just react to instructions, but can predict the consequences of their actions before performing them. Official CloudRobo samples: github.com/huaweicloud-samplโ€ฆ #Huawei #EmbodiedAI #WorldModels #Robotics
1
155