Almog Friedlander

Tech Talk: aleesha s j, our Senior Software Engineer, talks about quantization techniques in PyTorch to help optimize deep learning models for better performance and reduced memory usage. Learn the basics of quantization, PTQ vs QAT, symmetric vs asymmetric and per-tensor vs per-channel methods, the role of calibration, and see a hands-on demo including PyTorch quantization on DL model. Perfect for anyone working on edge-ready AI solutions!   Watch the video now: https://lnkd.in/gW2SpYDY   #TechTalk #PyTorch #Quantization #DeepLearning #ModelOptimization #EdgeAI #ONNX #MulticoreWare

34

The JEPA paper family, 𝘳𝘦𝘷𝘰𝘭𝘶𝘵𝘪𝘰𝘯𝘪𝘻𝘪𝘯𝘨 𝘈𝘐 𝘧𝘳𝘰𝘮 𝘵𝘰𝘬𝘦𝘯 𝘨𝘶𝘦𝘴𝘴𝘪𝘯𝘨 𝘵𝘰 𝘭𝘢𝘵𝘦𝘯𝘵 𝘳𝘦𝘢𝘴𝘰𝘯𝘪𝘯𝘨, really caught my attention and highlights a game-changing shift away from traditional LLMs’ purely generative paradigm. JEPA (Joint Embedding Predictive Architecture) originated in Yann LeCun’s 2022 position paper https://lnkd.in/dMUK5NbB and has since evolved through variants targeting self-supervised learning in vision, video, and multimodal domains. These works span foundational theory to specialized applications like LeJEPA and V-JEPA, all emphasizing latent-space prediction over token generation. By modeling semantic structure in latent space rather than superficial tokens or pixels, JEPA unlocks several key advantages:  -> Latent predictions, not raw reconstruction:     - Unlike SOTA LLMs glued to token-by-token generation (and prone to compounding errors), JEPA predicts compact embeddings of future or missing content—capturing meaning without being distracted by noisy surface forms.  -> Dramatic efficiency gains:     - Operating in a compact latent space (instead of huge token vocabularies) cuts parameter counts and compute while improving generalization—challenging today’s bloated LLMs on reasoning and multimodal tasks.  -> Superior multimodal fusion:     - Joint vision–language embeddings align modalities natively, excelling in cross-modal reasoning and transfer where token-based LLMs often struggle with disjointed vision-text handling.  -> Predictive stability over generative fragility:     - JEPA’s non-generative objective (predict structure, not raw pixels/text) leads to more stable training and richer world models—sidestepping classic LLM failure modes like brittle long-chain hallucinations. Recent benchmarks show VL-JEPA outperforms SOTA on visual reasoning, retrieval, captioning, and video prediction with leaner architectures. The future may be shaped by latent-first AI that truly understands before it generates. 2022 : Path Towards Autonomous... - Theory/world models https://lnkd.in/dMUK5NbB 2023: I-JEPA - Image-SSL arxiv.org/abs/2301.0824 2024-25: V-JEPA(2), ACT-JEPA - Video/Action arxiv.org/abs/2506.09985, arxiv.org/abs/2501.14622 2025-26: LeJEPA, VJEPA - Scal./Prob. arxiv.org/abs/2511.08544, arxiv.org/abs/2601.14354 #AI #JEPA #MultimodalAI #WorldModels #YannLeCu

6

AI pioneer Yann LeCun has joined the board of Nexar, an Israeli startup turning everyday cars into AI-powered sensors. This move signals a strategic shift toward real-world data for training AI. Nexar’s platform aggregates continuous video streams from vehicles worldwide, creating a massive dataset for building "world models" that can predict and reason in complex environments. LeCun, known for advocating physical-world grounding over pure LLMs, will guide this vision. The implication: safer, more reliable autonomous systems depend on vast, diverse real-world data, not just synthetic or text-based training. https://radu.ovh/7ouNNUV Stay ahead of AI innovations — subscribe to https://latent.press for daily curated news #AI #WorldModels #AutonomousVehicles #TechLeadership

2

1 Comment

𝗜𝗻 𝟮𝟬𝟮𝟲, 𝗣𝘆𝘁𝗵𝗼𝗻 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿𝘀 𝗮𝗿𝗲𝗻’𝘁 𝗷𝘂𝘀𝘁 𝗰𝗼𝗱𝗶𝗻𝗴, 𝘁𝗵𝗲𝘆’𝗿𝗲 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝘁𝗵𝗮𝘁 𝗱𝗼 𝘁𝗵𝗲 𝘄𝗼𝗿𝗸. Imagine one script that can: ✅ Reply to emails ✅ Analyze data ✅ Generate reports ✅ Run 24/7 Not sci-fi. This is Industrial Python in 2026. 🚀 𝗪𝗮𝗻𝘁 𝘁𝗼 𝗹𝗲𝗮𝗿𝗻 𝗣𝘆𝘁𝗵𝗼𝗻 + 𝗔𝗜 𝘀𝗸𝗶𝗹𝗹𝘀? 𝗖𝗼𝗺𝗺𝗲𝗻𝘁 “𝗟𝗶𝗻𝗸” #Python #AIAgents #Automation #DataEngineering #GenAI #CareerGrowth #growwithIdea

8

🌟 Recap — Cologne AI and Machine Learning Meetup #38 | Tomaz Bratanic “Agentic GraphRAG with MCP Servers” Tomaz (Neo4j) showed how graph memory turns agents from short-term chatbots into persistent, trustworthy assistants—and how MCP servers cleanly wire tools, data, and retrieval. Key takeaways 🧠 Graph memory as first-class context: entities & relations ground responses and reduce drift/hallucinations. ⚖️ Trade-offs that matter: graph + hybrid retrieval can boost accuracy; vectors alone often win on latency—pick per task, budget, and freshness. 🔒 Persistence & personalization: long-lived knowledge graphs enable user- and org-level memory with auditability and governance. 🙏 Thanks to lise GmbH for hosting and to everyone who joined the discussion! 📅 Follow us to get slides/updates and to be first in line for CAIML #39. #caiml #machinelearning #rag #graphrag #neo4j #knowledgegraphs #agents #mcp #llm #genai #community #cologne

25

1 Comment

🧑‍🏫 Dear Fellow Scientists, 🚀 Just Launched the "Course" for ML for Physicists Ep.2 ℹ️ If you’ve read the note and watched the video of episode 2 — you’ve seen how Loss Function connects with concepts in physics. 💻 Now, it’s time to get your hands dirty with code. We’ve prepared a Google Colab notebook that walks you through Loss Function the conventional way. 🔹 Run the code cells right inside your browser (no setup needed!) 📘 Table of Contents: * Loss Functions in Linear Regression * What is a Loss Function? * The Linear Model * MSE vs MAE * Setup and Imports * Load and Prepare the Dataset (California Housing) * Helper Functions: predict, MSE, MAE * Visualize Predictions and Residuals * MSE Loss Surface (w, b) * Gradient Descent for MSE (1D) * Results: Convergence and Fitted Line 👉 Take the Course: (Click the link and make sure to "Open with Google Colaboratory"): https://lnkd.in/gwm4448A 📖 The research note: https://lnkd.in/g8YXiVMK 🎥 The YouTube video: https://lnkd.in/ghuQPwJQ --- ℹ️ About these series (i.e. Notes, Courses, Projects): 🚀 Master Machine Learning with a Physics Mindset 📈 Get hired or promoted. Here’s your roadmap: 🔗 Go to program page: https://lnkd.in/gWwq78Vp 🎬 Start at Episode 1 and go step by step: 🔹 Read the Research Notes and watch the embedded videos to learn the concepts and how they connect with physics. 🔹 Take the Course: open the Colab notebook, run top to bottom, connect physics ideas to conventional ML. 🔹 Run the Project: apply concepts on real datasets to gain real-world experience. 🌟 Go further: turn your project into a professional GitHub repo and deploy a live web app. 📩 Click “Request Professional Setup” on the page to get started. 🎓 6-Month FREE-or-Land-Job Bootcamp From a math-based degree to a data science career: 💼 Land a Job or Pay Almost Nothing! 🧭 Step-by-Step Roadmap: 100% done-for-you structured schedule outlining exactly what to learn at each stage. 🧑‍🏫 Weekly Live Sessions: Expert-led project-solving sessions to master core data science skills through hands-on experience. 🌐 Build Your Online Presence & Portfolio: Work on real-world projects, receive expert reviews, and showcase your work on GitHub. 🧠 Technical & Soft Skills: Master all skills needed to land a job, from machine learning to interview and resume prep. 📄 Job Application Prep & Mentoring: Step-by-step guide on job search strategies, networking, and application materials. 🤝 Get Support When Needed: Connect with mentors for help, ask questions anytime, and join group discussions to learn and grow together. Apply now: 🔗 https://lnkd.in/gDv64aFg #Physics #MachineLearning #DataScience #AI

4

𝗔𝗰𝗰𝗼𝗿𝗱𝗶𝗻𝗴 𝘁𝗼 𝗚𝗮𝗿𝘁𝗻𝗲𝗿, 𝟲𝟬% 𝗼𝗳 𝗔𝗜 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝘄𝗶𝗹𝗹 𝗯𝗲 𝗮𝗯𝗮𝗻𝗱𝗼𝗻𝗲𝗱 𝗱𝘂𝗲 𝘁𝗼 𝗶𝗻𝗮𝗱𝗲𝗾𝘂𝗮𝘁𝗲 𝗱𝗮𝘁𝗮 𝗽𝗿𝗲𝗽𝗮𝗿𝗮𝘁𝗶𝗼𝗻. 𝗙𝗼𝗿 𝗮𝗴𝗿𝗶𝗰𝘂𝗹𝘁𝘂𝗿𝗮𝗹 𝗔𝗜, 𝘁𝗵𝗮𝘁 𝗻𝘂𝗺𝗯𝗲𝗿 𝗶𝘀 𝗹𝗶𝗸𝗲𝗹𝘆 𝗵𝗶𝗴𝗵𝗲𝗿. Field variability, inconsistent labeling, sparse edge case coverage are compounding data problems that cause models to fail between pilot and production. Here’s what actually kills agricultural computer vision projects: → Training on single-region datasets that don’t generalize across field conditions → Annotation inconsistency when labelers lack agricultural domain expertise → Missing edge cases (weather-damaged crops, early disease symptoms, unusual pest patterns) → Environmental variability outpacing data coverage Standard ML pipelines aren’t built to handle the biological complexity and environmental chaos of real-world agriculture. Here are four data challenges you should be aware of if you don’t want your agricultural AI stuck in pilot 👇 #AgTech #MachineLearning #ComputerVision #AI #PrecisionAgriculture #DataScience

8

🎙️ 𝗘𝗽𝗶𝘀𝗼𝗱𝗲 #𝟱𝟬 – 𝗔 𝗖𝗮𝗻𝗱𝗶𝗱 𝗖𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻 𝗼𝗻 𝗗𝗮𝘁𝗮 𝗣𝗿𝗼𝗱𝘂𝗰𝘁𝘀, 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜, 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗘𝘃𝗼𝗹𝘃𝗶𝗻𝗴 𝗥𝗼𝗹𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗖𝗗𝗢 𝗳𝗲𝗮𝘁𝘂𝗿𝗶𝗻𝗴 Veronika Durgin 👉 𝗨𝗥𝗟: https://lnkd.in/gk_7uNNk In this milestone episode, Veronika Durgin joins the conversation to unpack critical shifts shaping the data landscape — from how organizations define and govern Data Products to the emergence of Agentic AI and the balancing act between centralized and federated approaches. 𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀: ✓ 𝗗𝗮𝘁𝗮 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝗧𝗵𝗶𝗻𝗸𝗶𝗻𝗴 – Why clarity of ownership, service-level expectations, and discoverability is vital. ✓ 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 – What it means for data teams and how it intersects with automation, orchestration, and adaptive decision-making. ✓ 𝗥𝗼𝗹𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗖𝗗𝗢 – How modern data leaders must bridge strategy, execution, and organizational design. ✓ 𝗙𝗲𝗱𝗲𝗿𝗮𝘁𝗲𝗱 𝘃𝘀 𝗖𝗲𝗻𝘁𝗿𝗮𝗹𝗶𝘇𝗲𝗱 – Lessons in designing scalable operating models and how the DPM team can become a true enabler, not just a gatekeeper. This isn’t just a technical discussion — it’s a strategic look at how data leaders must evolve in the age of intelligent platforms and distributed accountability. 👉 𝗦𝘂𝗯𝘀𝗰𝗿𝗶𝗯𝗲 𝘁𝗼 𝗟𝗲𝘁’𝘀 𝗧𝗮𝗹𝗸 𝗔𝗯𝗼𝘂𝘁 𝗗𝗮𝘁𝗮! 𝗳𝗼𝗿 𝗺𝗼𝗿𝗲 𝗶𝗻-𝗱𝗲𝗽𝘁𝗵 𝗲𝘅𝗽𝗲𝗿𝘁 𝗶𝗻𝘀𝗶𝗴𝗵𝘁𝘀: [https://lnkd.in/gkz_FK8f] #DataProducts #AgenticAI #ChiefDataOfficer #FederatedData #DataStrategy #LetsTalkAboutData #Podcast

1

🧩 𝗗𝗮𝗴𝗴𝗿 — Visual AI Workflows in Pure Python (by Gradio) #viix #AI #ml_ops #python #Gradio #ml_workflows #Daggr If you’ve ever built ML pipelines with scripts and wished for better visibility — but didn’t want to jump into heavyweight orchestration platforms — Daggr might be exactly what you need. Developed by the team behind Gradio, Daggr is an open-source Python library that lets you build AI workflows connecting: 🔸 Gradio apps 🔸 ML models via Hugging Face Inference Providers 🔸 Custom Python functions And here’s the key part: it automatically generates a visual graph (canvas) of your workflow — ComfyUI-style, but Python-native. 🚀 What Makes Daggr Interesting for Developers? ✅ Code-First Approach You write standard Python logic. Daggr turns it into an interactive DAG automatically — no manual graph wiring required. ✅ Hot Debugging (Game Changer) If something breaks mid-pipeline, you don’t need to restart everything. You can: 🔶 Inspect intermediate outputs 🔶 Modify parameters 🔶 Rerun a single node 🔶 Continue from that point For long-running or multi-step workflows, this saves serious time. ✅ State Preservation Daggr maintains process state across runs — ideal for iterative experimentation and complex chains. ✅ Provenance Tracking When reviewing past results, Daggr restores the exact inputs that produced them. It also highlights which parts of your workflow are stale after changes. 🛠 Where It Fits in the Ecosystem Daggr sits in a practical middle ground: 🔶 More structured and inspectable than plain scripts 🔶 Much lighter than enterprise orchestration platforms 🔶 More Pythonic than purely visual tools It integrates naturally with: 🔶 Gradio Spaces 🔶 Hugging Face 🔶 Remote inference providers Example workflows are already available on Hugging Face Spaces, making it easy to explore real-world setups. 💡 Why It Matters As AI systems become more modular (LLMs + tools + retrieval + vision + custom logic), debugging and observability become critical. Daggr gives developers: ✔️ Workflow transparency ✔️ Fast iteration cycles ✔️ Reproducibility ✔️ Visual clarity without sacrificing Python control It feels like bringing the ergonomics of node-based UI systems into a clean, developer-first ML workflow environment. For anyone building demos, research pipelines, or experimental AI products — this is definitely worth exploring. 🔗 GitHub: https://lnkd.in/g37RaiFr 🤗 HF: https://lnkd.in/d5tcrCSw

5

Major Performance Breakthrough in Gradio 5.36 We just solved a BIG performance bottleneck in ML apps: > Previous versions rendered all components regardless of visibility, creating significant performance costs. > The new selective rendering system only processes visible components, delivering: - dramatic reduction in memory consumption for complex apps - faster load times for multi-tab UIs Ready to upgrade? pip install --upgrade gradio What additional performance improvements would you like to see in your ML apps? Let us know in comments!

138

3 Comments

#story AndRej Karpathy built an AI that does autonomous ML research. - picks architectures - tunes hyperparameters - commits code changes. - runs experiments in a loop - zero human input -/https://lnkd.in/dCayaAB2

1

🚀 Unlocking the Power of Amazon S3 Vectors in Page Initialization As a consultant constantly exploring better ways to optimize data workflows and boost performance, I’m excited to share a new resource we’ve just launched: a dedicated page on S3 Vectors for Page Init. Why does this matter? When building scalable, data-heavy applications, how and where you store and retrieve vector embeddings can make or break your speed, cost, and user experience. With Amazon S3 as a vector store, combined with efficient page initialization strategies, we’re looking at: ✅ Reduced cold start times ✅ Massively scalable, low-latency retrieval ✅ Simplified infra with fewer moving parts This new page dives into the how, the why, and the when of using S3 vectors to streamline initialization — from ML-driven personalization to real-time content rendering. 🧠 Whether you're working on AI applications, semantic search, or just trying to optimize your load times, there’s something here for you. 🔗 Check it out, and let’s discuss — how are you using S3 for modern vector workflows? #AWS #S3Vectors #VectorDatabase #PageInit #AIInfrastructure #ScalableArchitecture #ConsultingLife #TechInnovation #MachineLearning #StartupTech #DataOptimization

6