Weixuan XIAO

New video on distributed training is out ! This time you will learn the theory behind Tensor Parallel and how to implement it yourself in Picotron ! Link of the video in the comment section !

183

7 Comments

We just released LangGraph v0.6 over at LangChain! This is our last minor release before the big v1 milestone 🎯. ✨ A new context API for cleaner, type-safe runtime dependency injection 🔀 Dynamic model & tool selection for create_react_agent 🛡️ Enhanced type safety & autocomplete for graph building and invocation  🏗️ durability mode for fine-grained persistence control See our release notes for a full summary + some short demo snippets: https://lnkd.in/eWzB729F.

722

26 Comments

Hi Network, I ported #Mosquitto to ESP-IDF -- to see if it could actually run on an ESP32. Here’s what I learned 👇 Blog post: 𝗟𝗶𝗴𝗵𝘁𝘄𝗲𝗶𝗴𝗵𝘁 𝗠𝗤𝗧𝗧 𝗯𝗿𝗼𝗸𝗲𝗿 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗼𝗻 𝗘𝗦𝗣𝟯𝟮 https://lnkd.in/ey3YYQvu 🔍 Feasibility – does it even make sense? 🎯 Use cases – when would you want this? 🛠️ Technical details – memory, threading, what I had to trim 𝘈𝘵𝘵𝘳𝘪𝘣𝘶𝘵𝘪𝘰𝘯 𝘯𝘰𝘵𝘪𝘤𝘦:🎥 Generated with PixVerse model pixverse-v4.5.

82

5 Comments

🔧 𝗛𝗼𝘄 𝗰𝗮𝗻 𝘁𝗵𝗲 𝗔𝗿𝗰𝗮𝗱𝗶𝗮 𝗺𝗲𝘁𝗵𝗼𝗱 𝗯𝗲 𝘂𝘀𝗲𝗱 𝘄𝗶𝘁𝗵 𝗦𝘆𝘀𝗠𝗟 𝘃𝟮? We've explored this question by extending SysON, our open-source modeling solution, to support Arcadia concepts. This first experiment is now available in a blog post, with examples, diagrams, and early feedback on the approach. 👉 https://lnkd.in/eRvJ25i7 #MBSE #SysMLv2 #Arcadia #SysON #OpenSource #SystemsEngineering

58

1 Comment

Mistral AI released Mistral OCR that understands documents and can comprehend text with tables, equations, media → parses 1000 pages for just $1 and 2000 if using batch inference → state of the art in understanding complex layouts and document with images, math expressions → outperforms google document ai, azure ocr, and other models like gemini 2.0 flash on multiple benchmarks → understands multiple languages (works amazing well on hindi) → can process upto 2000 pages per minutes on a single node → get output in structured format like json → you can also self-host. explore this by talking to the mistral team i tried it on few cases and it worked really well. i even tried it on sanskrit and got fairly decent results blog post 📄 https://lnkd.in/g48emUWt try 🕶️ it on le chat https://lnkd.in/gdz2E984 cookbook 👨‍🍳 https://lnkd.in/gcpAmz5Y

216

5 Comments

Samsung’s Tiny Recursive Model (TRM) is a reminder that architecture and compute strategy can matter as much as parameter count. What’s new • 7M params, trained on ~1k examples, uses a draft → latent scratchpad → self-critique → revise loop (up to 16 cycles). • Reports 45% on ARC-AGI-1 and 8% on ARC-AGI-2, exceeding many much larger LLMs. • Code + paper are public:   · Blog → https://lnkd.in/d5vriPDc  · Paper → https://lnkd.in/eJTUYtFC  · Code → https://lnkd.in/eEmNc9Jj Why it’s interesting • Shows a recursive think-then-rewrite controller can boost reasoning without billion-scale models. • Validates a neuro-symbolic flavor (explicit inner loop, scratchpad, critique). • Points to cheap, deployable reasoners for constrained tasks at the edge. Nuance (observer lens) • TRM is specialized (e.g., Sudoku/Maze/ARC) and its ARC results rely on heavy augmentation, recursion depth, and voting; so compare by compute budget (FLOPs/latency), not just parameters. • Great for structured puzzle-like problems; not a general LLM replacement. The right takeaway: algorithmic loops + small nets can win on certain classes of reasoning. What we’re watching at RediMinds • Where recursive controllers fit into production stacks (agent toolflow → deliberate inner loops → verifiable outputs). • Eval design that normalizes for compute, not only params. • Opportunities for edge inference in regulated and resource-constrained settings (gov/legal/health ops) where small, auditable models shine. #AIReasoning #AlgorithmicEfficiency #EdgeAI #LLMops #AIEngineering #RediMinds #CreateTheFuture

88

Software relies on an operating system, and software engineers can construct better software if they understand how an operating system works. If you want to learn about operating systems without plodding through millions of lines of code, take a look at the third edition of Operating System Design -- The Xinu Approach, available now (ISBN 9781032980997). It covers all all the functional pieces and explains how an operating system manages concurrent processes, I/O devices, memory, and network communication. New chapters cover pipe mechanisms and multicore systems.

214

10 Comments

When developing detection rules, what defines a "great detection" depends entirely on who's using it. Nasreddine Bencherchali illustrates this perfectly in his new article on detection engineering. The same detection can simultaneously be: • Brilliant for a DFIR investigator looking for suspicious threads during an incident • A noisy nightmare for an MSSP/SOC analyst processing millions of daily events • A performance liability for detection platform engineers Key takeaways: • Understand your audience's operational reality before writing rules • Include clear metadata that sets appropriate expectations • Consider available telemetry when designing detections • A detection without a defined audience serves nobody well Detection engineers should ask: Who am I writing this for? What's their operational reality? What does success look like for them? How easily can they operationalize this? 🚀 Turn theory into practice. Learn actionable detection engineering and AI agents for security you can implement immediately. Secure your spot: https://buff.ly/CXPknu4

85

Yet Another New Multimodal Fine-Tuning Recipe �� 🧑‍🍳 In this Hugging Face Cookbook notebook, we demonstrate how to align a multimodal model (VLM) using Mixed Preference Optimization (MPO) with trl. 💡 This recipe is powered by the new MPO support in trl, enabled through a recent upgrade to the DPO trainer! We align the multimodal model using multiple optimization objectives (losses), guided by a preference dataset (chosen vs. rejected multimodal pairs) Check it out! ➡️ https://lnkd.in/dUsxGtpH

238

5 Comments

⚡️Larger instance sizes: 48 & 64 CPU TigerData (creators of TimescaleDB) Cloud now supports our biggest dedicated instances yet: ・48 CPU, 192 GB RAM ・64 CPU, 256 GB RAM Pair with up to 10 auto-load-balanced replicas → scale to 704 cores & 2816 GB RAM for your most demanding applications. Then extend with: 📦 Data Tiering → petabytes of S3-backed storage 🌊 Tiger Lake → continuous replication to your lakehouse Postgres without Limits. Not launch week, just shipping daily. 3/3 today. 💥 #ProductionGrade #BuiltTigerTough

53

Hallucinations + Infinite Loops: The Reliability Bottleneck in Multi-Agent AI When experimenting with multi-agent orchestration (LangChain + AutoGen), I kept running into the same two failure modes: 🔹 Hallucination Propagation One agent fabricates an API response → the next accepts it as ground truth → errors compound. Example: A “research agent” invents a hotel name, and a “booking agent” tries to reserve it via API. 🔹 Runaway Loops Recursive delegation without clear termination. Example: Two planning agents bouncing the same task back and forth for 40+ turns. Both problems waste tokens, spike latency, and erode trust. What worked in practice: 1. Structured Outputs as Contracts → JSON/Pydantic validation before handoff. 2. Verifier Agents → lightweight gatekeepers to sanity-check citations, API calls, or calculations. 3. Loop Detection → max_turn thresholds + state hashing to kill cycles early. 4. Bounded Reflection → constrain “improve answer” to N iterations instead of chasing perfection. The takeaway: Multi-agent AI doesn’t fail because the models are weak — it fails because the system lacks guardrails. The future of agentic AI isn’t “more autonomy,” it’s engineered autonomy: autonomy with contracts, supervisors, and verifiers baked in. Curious — if you’ve worked with multi-agent frameworks, how are you approaching reliability at scale? Are you building custom watchdogs, or relying on framework defaults?

46

5 Comments

LangChain just dropped Open Agent Platform — an open-source, no-code platform for building AI agents. It’s designed for non-developers, but powerful enough for advanced use cases. You can build customizable agents that connect with: 🛠️ MCP tools 📄 LangConnect for RAG 🤖 Other LangGraph agents Also comes with two plug-and-play agents: • MCP/RAG Tools Agent • Multi-Agent Supervisor A big step toward making agents more accessible and interoperable. Link : https://lnkd.in/gvpW3cUR https://lnkd.in/g9-v5RJn

83

2 Comments

🧠 "I want to learn Linux… but where do I even start?" Here’s a carousel I wish someone gave me when I started working on HPC. During my 3 years of PhD at CEA Marcoule, France 🇫🇷, I worked extensively on high-performance computing systems. I boiled it down to a bare-minimum set of commands you actually need — no fluff, just what makes you to work with terminal. This post covers: 1️⃣ Minimal Linux commands to navigate files and folders (pwd, ls, mkdir, cd, mv, cp…) 2️⃣ Commands to transfer data between your local computer and a remote HPC/Cloud server using ssh & sftp 👉 If you just master these few commands — and use them consistently — you’ll be able to perform 90% of your day-to-day work on an HPC or remote Linux server without needing any file manager. Yes, everything from the terminal. 💡 Note: There are always multiple ways to do the same thing on Linux. But this is the most beginner-friendly path that I’ve found works reliably for newcomers to HPC environments. 🧰 In the next post, I’ll cover: 1️⃣ Editing files via vi/vim editor 2️⃣ Submitting and managing jobs on HPC (sbatch, squeue, scancel, sacct…) Save this post. Share it with your labmate who’s just getting started. #HPC #Linux #CommandLine #PhDLife #ScientificComputing #Bioinformatics #TerminalCommands #RemoteServer #Ubuntu #LinuxBasics #CEAMarcoule #ScientiFlow #HighPerformanceComputing #AcademicLife #LinuxForScientists #HPCTraining #BashBeginners

46

4 Comments

📢 Big developments in the world of #GaussianSplatting: 𝟯𝗗𝗚𝗥𝗨𝗧 by NVIDIA Research: a hybrid approach for implementing 𝟯𝗗 𝗚𝗮𝘂𝘀𝘀𝗶𝗮�� 𝗥𝗮𝘆 𝗧𝗿𝗮𝗰𝗶𝗻𝗴 (𝟯𝗗𝗚𝗥𝗧) and 𝟯𝗗 𝗚𝗮𝘂𝘀𝘀𝗶𝗮𝗻 𝗨𝗻𝘀𝗰𝗲𝗻𝘁𝗲𝗱 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺 (𝟯𝗗𝗚𝗨𝗧) 📖 𝗧𝗟𝗗𝗥: the paper extends 3D Gaussian Splatting (3DGS) to support fisheye or rolling-shutter cameras, the latter being commonly used in autonomous driving and robotics. The solution also introduces secondary rays for simulating reflections and refractions. 💡𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀: - it's now easier to create Gaussian splats by using 360 cameras (fewer photos required), or even using footage from autonomous cars. - any synthetic objects you add to a scene can now reflect and refract light without using a more performance intensive method like ray tracing. Creating realistic experiences on top of your Gaussians just became a lot easier! 👇 Link to the paper is in the comments. Michael Rubloff Michal Gula Cody A. Sherrill Sebastian G. have you played around with the project?

49

15 Comments