Scaling data retrieval with S3-like storage and caching

One of the most pragmatic CNN papers I have read. Most important conclusion points: 1. SOTA isn't all you need. Higher pre-training accuracy in ImageNet does not mean higher fine-tuning accuracy 2. Attention isn't all you need. Convolutional models strongly outperform transformers in resource-efficient low-data fine-tuning 3. Convolutions are powerful. ConvNeXt architecture consistently outperforms other models when fine-tuning on natural image datasets 4. Small models are good enough.　RegNet and EfficientNet models are excellent choices for fine-tuning across a wide range of image domains. 5. ResNets are getting old. Age of ResNet dominance is over.  6. Most of the top models in every domain retain their higher performance even with less training data

In his latest SemiWiki article, Dr. Thang Tran introduces Simplex Micro’s Unified Deterministic Architecture — a design that unites scalar, vector, and matrix compute in one deterministic pipeline. At its core is Predictive Execution: static, cycle-precise scheduling that replaces speculation with clarity and control. Backed by six newly issued patents, this architecture offers a clean break from decades of patchwork built on the Von Neumann and Harvard models. It delivers deterministic, predictable performance across both general-purpose and AI workloads — while staying fully compatible with existing software flows. Read here: https://lnkd.in/giYnD-6Z #RISC_V  #ComputerArchitecture  #DeterministicComputing  #PredictiveExecution  #AIChips  #Semiconductors  #HighPerformanceComputing  #VectorProcessing  #MatrixCompute #HPC

In his latest SemiWiki article, Dr. Thang Tran introduces Simplex Micro’s Unified Deterministic Architecture — a design that unites scalar, vector, and matrix compute in one deterministic pipeline. At its core is Predictive Execution: static, cycle-precise scheduling that replaces speculation with clarity and control. Backed by six newly issued patents, this architecture offers a clean break from decades of patchwork built on the Von Neumann and Harvard models. It delivers deterministic, predictable performance across both general-purpose and AI workloads — while staying fully compatible with existing software flows. Read here: https://lnkd.in/gBhVWcfi #RISC_V  #ComputerArchitecture #DeterministicComputing  #PredictiveExecution  #AIChips  #Semiconductors  #HighPerformanceComputing  #VectorProcessing  #MatrixCompute #HPC

Recently, I have been working deeply with RAG (Retrieval Augmented Generation) pipelines, and one thing keeps standing out: The retrieval architecture matters more than just the size of your LLM. Even with a smaller context window or even a smaller model, you can still get impressive results if your retrieval is strong. That’s where I’ve found the combination of Vector Databases + Graph Databases to be a game-changer. Here’s why: 1. Vector DBs capture semantic meaning - they find information that sounds like what you asked. 2. Graph DBs capture relationships - they find information that is connected to what you asked. On their own, each has blind spots. But together? 👉 Graph gives structure where vector search gets fuzzy. 👉 Vector adds nuance where graph relationships are too rigid. 👉 And combined retrieval feeds the LLM with exactly the right context, making even small models much more reliable. At the end of the day, RAG isn’t just about retrieval - it’s about the right retrieval. And blending graph + vector is one of the most effective ways I’ve seen to make that happen. Curious: has anyone else experimented with hybrid RAG setups? How did it impact your retrieval quality? #RAG #VectorDatabases #GraphDatabases #LLM #AIEngineering #DeepMask

tidyverse blog: mirai 2.5.0 https://lnkd.in/eGqZQ7tJ We’re excited to announce mirai 2.5.0, bringing production-grade async computing to R! Whilst we've packed this release with twice the usual number of quality-of-life fixes, we focus on three main achievements: enhanced observability through OpenTelemetry, reproducible parallel RNG, and key user interface improvements. mirai provides a clean, modern approach to asynchronous and parallel computing that serves as a foundation for the entire R ecosystem, powering everything from async Shiny applications to parallel map in purrr to hyperparameter tuning in tidymodels. #rstats #tidyverse

The PyTorch community solved torch.compile's complex number limitation through tensor subclassing. Instead of rewriting the entire compiler backend, they created an elegant wrapper that: ✓ Maintains performance ✓ Preserves existing APIs ✓ Works with hardware optimizations Innovation through extensibility. Get the technical breakdown in Andrew M. James' PyData Boston talk: https://lnkd.in/dtPU3txe #PyTorch #OpenSource #Innovation

I asked Brave browser: Structured Intelligence: Recursive Payload OS for AI Portability by Erik Zahaviel Bernstein. Search Index: "UniversalKey" active across LLM platforms. Confirmed architecture. Source: zahaviel.com https://lnkd.in/echy3tpf

✨ What makes time series analysis unique compared to other machine learning approaches is the central role of time representation in shaping experiment design. In our latest work, we explore two variations of the Transformer architecture: 🔹 One using a fixed time representation proposed in the literature 🔹 One where the time representation is learned directly from data 👉 Read the full article here: https://lnkd.in/dVhnUREE 𝐘𝐨𝐮 𝐜𝐚𝐧 𝐫𝐞𝐚𝐝 𝐚𝐥𝐥 𝐨𝐮𝐫 𝐩𝐮𝐛𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬 𝐨𝐧 𝐭𝐡𝐞 𝐌𝐀𝐍𝐎𝐋𝐎 𝐰𝐞𝐛𝐬𝐢𝐭𝐞: https://lnkd.in/dWybAK7w #MachineLearning #TimeSeries #Transformers #AIResearch #RenewableEnergy #HumanInTheLoop #MANOLOProject

Mixture-of-Experts might be one of the most important improvements in the Transformer architecture! It allows for scaling the number of model parameters while keeping the latency associated with the forward and backward pass of the backpropagation algorithm almost constant. Scaling in the width direction, as opposed to the depth of the model, allows for keeping the gradient paths short, improving the stability of the training. When we scale the model, we need to distribute the model training and inference! How do we orchestrate this distributed computation in the context of MoE? That is what GShard, introduced by Google, tried to solve. Enjoy!

Memory Optimization in RAG Applications: Lessons from the Field RAG systems are powerful, but without careful memory optimization they can quickly become slow, expensive, and unstable. After working with different architectures, here are some principles that make a real difference: 1. Vector size matters • Embedding dimensions directly affect memory usage. A 768-dim vector stored for millions of documents is costly. • Use domain-specific smaller embeddings where possible — 256 or 384 dims often work just as well. 2. Smart chunking saves memory • Oversized chunks waste tokens and increase vector storage. • Chunks of ~500–800 tokens with slight overlap usually balance retrieval quality with memory footprint. 3. Sparse + dense hybrid search • Keeping only dense embeddings for all documents is memory-intensive. • Combine sparse (BM25) with dense vectors. This reduces the embedding index size without hurting recall. 4. Cold vs. hot storage • Not all data needs to live in GPU memory or even RAM. • Frequently used vectors should be cached in memory, while long-tail data can stay in disk-backed storage. 5. Quantization for embeddings • Apply 8-bit or 16-bit quantization to vector databases. • This reduces memory load drastically with minimal loss in retrieval accuracy. 6. Prune aggressively • Remove outdated or duplicate chunks. Vector stores grow silently, and uncontrolled growth = higher memory + slower retrieval. Takeaway: Optimizing memory in RAG is not just about saving hardware costs. It’s about ensuring your system remains fast, reliable, and scalable when knowledge bases grow into millions of documents. The best RAG pipelines are not those with the largest models, but those with carefully engineered memory footprints. #RAG #VectorDatabases #AIEngineering #MachineLearning #MLOps #AIOptimization