José Manuel de la Chica’s Post

- DeepSeek V3.2-Exp debuts DeepSeek Sparse Attention （DSA）: a lightning indexer + fine-grained top‑K token selection that cuts attention complexity from O（L²） to O（Lk） while preserving quality. - Real-world efficiency: 50% API price cut; 2–3x faster long‑context inference; 30–40% lower memory; 50% faster training; 10x cheaper 128K token decoding vs dense. - Rigorous training pipeline: dense warm‑up then sparse training （943.7B tokens）, followed by post‑training with specialist distillation and single‑stage GRPO for balanced rewards. - Hardware co-design: optimized FlashMLA/DeepGEMM kernels on Hopper H800 GPUs hit near-peak bandwidth （3000 GB/s）, with open-source PRs. - Performance profile: parity overall vs V3.1; clear gains in coding/agentic tasks; minor regressions on ultra-abstract reasoning benchmarks （e.g., GPQA Diamond, HMMT）. 🔔 Follow us for daily AI updates! 📘 Facebook: https://lnkd.in/gxDt7PJa 📸 Instagram: https://lnkd.in/gmYfWDbF #DeepSeek #AI #LLM #GenerativeAI #AIGenerated #CreatedWithAI

Why AI Fails: Data Ingestion Patterns Matter More Than Models Most AI failures don’t stem from weak models—they often begin with flawed data ingestion. This essay examines key design patterns, including Full Loader, Incremental Loader, and CDC, demonstrating how they influence trust, scalability, and performance. If you build or run ML systems, these insights will help you strengthen the foundation on which your models depend. Curious which ingestion pattern protects your ML pipeline from silent failure? Dive into the full breakdown here. #DataIngestion #MachineLearningOps #ChangeDataCapture #AIinProduction #StreamingAnalytics #BigDataEngineering #PredictiveMaintenance #RealTimeAI #DataPipelines #AITrust

I literally loved the straight forward explanation in the docs about the actual architecture of how Mem0 adds up the important stuff into the memory. Simply put, Mem0 actually runs a small internal LLM pipeline before persisting the data. Basically, it’s distilling semantic essence from your text - kind of like how we humans form “mental summaries” when remembering something. Then these extracted “facts” (which are cleaner, smaller, and more meaningful) get stored as memory embeddings + structured metadata inside the Mem0 store. It’s not just retrieval - it’s understanding before storing ❤️🔥 #Mem0 #AI #LLM #AIEngineering #ArtificialIntelligence #Memory #LangGraph #LangChain #GenerativeAI #LearningInPublic #MachineLearning

State of LLMs in Late 2025 The AI landscape has evolved into a hyper-specialized ecosystem where each model has distinct strengths. The challenges emerging include diminishing returns on scaling, massive energy consumption, and the rise of smaller specialized models. This guide explains the technical foundations that make each model different so developers can choose the right tool for their task. Success now means understanding each model's strengths, testing on specific use cases, routing intelligently, and staying current. https://lnkd.in/ednh6N9v

Why do multi-agent AI systems collapse at scale? It’s not communication gaps, but missing memory. Without shared, persistent memory, agents repeat work, lose context, and drive up token costs exponentially. This article explores memory engineering—from structured memory blocks to a MongoDB Atlas–powered exocortex, that aligns agent states, optimizes data retrieval, and scales coordinated AI teams with reliability and cost efficiency. Learn more: https://lnkd.in/g2uB5ggA

𝗗𝗲𝗲𝗽𝘀𝗲𝗲𝗸𝘀 𝗻𝗲𝘄 𝗺𝗼𝗱𝗲𝗹 𝘃𝟯.𝟮-𝗲𝘅𝗽 has landed. The most important feature of the new experimental model is called DeepSeek Sparse Attention, the architecture allows the models to operate over long portions of context with comparatively small server loads. This means the price of a simple API call could be reduced by as much as half in long-context situations.  Great for chatbots but because of the nature of processing (scans long sentences for key words, as opposed to word pairs) it may lack accuracy needed for medical or legal requirements. Test away, its on Hugging Face #it #ai #compute https://lnkd.in/g64u7Gy7

Qwen3‑Next: Redefining Efficiency in Large Language Models The Qwen team just unveiled Qwen3‑Next, a major leap in model architecture — balancing scale, efficiency, and long‑context reasoning like never before. Core breakthrough: - 80B parameters, but activates only ~3B per inference step - Matches the dense Qwen3‑32B performance - Trains at < 10% of the cost - Delivers > 10× faster inference on 32K–256K token contexts Key innovations: - Hybrid Attention (Gated DeltaNet + Gated Attention) – combines the recall of standard attention with the efficiency of linear attention (optimal 3 : 1 mix). - Ultra‑Sparse MoE – 512 experts with only 3.7 % active at a time, dynamically balanced for stability and throughput. - Stability‑Optimized Training – Zero‑Centered  RMSNorm and normalized routing for reliable convergence under high sparsity. - Multi‑Token Prediction (MTP) – enables efficient speculative decoding and higher generation speeds. Open‑source availability: On Hugging Face, ModelScope, and compatible with vLLM + SGLang for up to 256K context. The new Instruct and Thinking models outperform their 30B–32B counterparts and approach results of massive 235B models — a compelling case for smarter, not bigger, model scaling. Takeaway: The next era of LLMs is defined by two fronts — context‑length scaling and activation sparsity. Qwen3‑Next shows that real progress now lies in efficiency engineering, not raw parameter count. Full deep dive https://lnkd.in/eEuY-7-4 #AI #LLM #Qwen #MixtureOfExperts #DeepLearning #GenerativeAI #ModelOptimization #AIResearch #Efficiency

𝗘𝘃𝗲𝗿𝘆 𝗺𝗶𝗹𝗹𝗶𝘀𝗲𝗰𝗼𝗻𝗱 𝘆𝗼𝘂 𝘀𝗮𝘃𝗲 𝗶𝗻 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗺𝗶𝗴𝗵𝘁 𝗱𝗲𝗽𝗲𝗻𝗱 𝗼𝗻 𝗼𝗻𝗲 𝘁𝗵𝗶𝗻𝗴: 𝗵𝗼𝘄 𝘆𝗼𝘂 𝗵𝗮𝗻𝗱𝗹𝗲 𝘁𝗵𝗲 𝗞𝗩 𝗰𝗮𝗰𝗵𝗲. One of the main challenges in serving LLMs efficiently is managing what’s called the 𝗞𝗩 𝗰𝗮𝗰𝗵𝗲 (Key-Value attention cache). When you send a prompt to an LLM, the model doesn’t just process it once and move on. For every token it generates, it needs to look back at all the previous tokens (your query and the tokens it already produced) to decide what comes next. To avoid recomputing that entire history every time, the model stores the key and value representations of past tokens. This is the KV cache. It sounds simple if you’re familiar with transformer architectures, but this cache can grow massive, sometimes even larger than the LLM itself. And since it lives in memory, managing it efficiently becomes one of the biggest bottlenecks in LLM inference. And when most of the compute bill today comes from inference, every improvement in KV cache handling matters. In future posts, we’ll dive deeper into how different serving frameworks handle this challenge, because managing memory is just as critical as managing models. #LLM #AI #MachineLearning #DeepLearning #MLOps #AIInfrastructure #KVCaching #LLMServing #Inference

Why do multi-agent AI systems collapse at scale? It’s not communication gaps, but missing memory. Without shared, persistent memory, agents repeat work, lose context, and drive up token costs exponentially. This article explores memory engineering, from structured memory blocks to a MongoDB Atlas–powered exocortex, that aligns agent states, optimizes data retrieval, and scales coordinated AI teams with reliability and cost efficiency. Learn more: https://lnkd.in/gyDN4ytG

Why do multi-agent AI systems collapse at scale? It’s not communication gaps, but missing memory. Without shared, persistent memory, agents repeat work, lose context, and drive up token costs exponentially. This article explores memory engineering—from structured memory blocks to a MongoDB Atlas–powered exocortex, that aligns agent states, optimizes data retrieval, and scales coordinated AI teams with reliability and cost efficiency. Learn more: https://lnkd.in/gCqKib9a