Rudy Pei

⚙️ One library, real DSP speed. There’s a gap in the developer ecosystem. Teams doing DSP/ML often migrate to Python, because that’s where most tooling lives. But Python wasn’t built for high-throughput I/O — under load, it spawns new interpreters/processes, creating duplicated memory, slow cold starts, and more containers. Compute stays fast, but orchestration cost explodes. Meanwhile, Node.js/TS scale I/O well, but were never meant for serious compute. Pure-JS DSP hits performance walls. Even worse, serializing data between Redis ↔ DSP/ML adds latency and forces batching — slowing real-time pipelines. So devs are stuck: Python → strong math, weak concurrency Node → strong networking, weak compute dspx closes that gap — TypeScript DX + native C++/SIMD (AVX2/SSE3/SSE2/NEON) under the hood. 🧩 How dspx works FIR / FFT / Conv1D run in optimized C++/SIMD Memory-safe circular buffers → O(1) throughput TS handles Kafka / Redis / WebSockets Redis persistence < 0.5 ms Batched logging avoids I/O stalls At tiny batch sizes, N-API overhead can make naive JS look competitive. At realistic scale, batched native pipelines win. ⚡ Benchmarks Dell OptiPlex 3000 Micro (i5-12600T · AVX2 · Node 22) FFT: 2.6× faster than fft.js, ~9× faster than tfjs-node FIR (51-tap): ~4× faster than fili / naive JS Conv1D (128-kernel): ~3.5× faster than tfjs-node Moving Avg (O(1)): 559× more throughput/sec than naive JS Redis save/load: sub-ms Logging: <3% overhead 📊 Full tables + charts in carousel 💬 Observations SIMD FMA + contiguous memory → big FIR/Conv wins O(1) moving-average design → massive throughput gain Sub-ms Redis + low I/O overhead → real-time persistence in pure Node 🧠 Open Source dspx is Apache 2.0, free for commercial/academic use. Looking for contributors interested in: ARM NEON / Apple M / Graviton tuning Audio / sensor / biomedical DSP Visualization + benchmark tooling 📦 npm i dspx 🔗 https://lnkd.in/e-tWxAgu 🧠 Real-time DSP for Node, TypeScript & Redis 💼 Note I’m exploring opportunities in full-stack, real-time systems, performance engineering, and DSP. If your team works in this space, I’d love to connect. My next post will cover why sub-ms Redis latency isn’t just technical — it’s economic. Lower serialization + compute overhead → lower infra + energy cost. 🔖 Tags & Mentions #NodeJS #TypeScript #DSP #PerformanceEngineering #EdgeComputing #OpenSource #SIMD #Redis #Cplusplus #RealTime NodeJS Developer TensorFlow Google Microsoft JavaScript Developer Amazon Web Services (AWS) Vercel

5

1 Comment

For a decade, Yann LeCun worked at FAIR (Facebook AI Research). He thrived under a philosophy of open science and academic freedom. He pushed to release Meta’s Llama AI models to the public, aiming to democratize access to powerful artificial intelligence tools. But in the past few years, things started to change, as Meta focused on directly competing with big AI companies. The breaking point arrived in 2025. Meta reorganized its AI efforts under Alexandr Wang, the 29-year-old founder of Scale AI. LeCun seemed to view Wang as a scaling maximalist — someone who believed that if you just throw enough GPUs and data at a model, it will eventually wake up and become smart. LeCun disagreed. He argued that you could scale an LLM until it consumed the energy of the sun, and it would still struggle to truly understand why a glass shatters when it hits a floor. https://lnkd.in/g-d-jf-t #FatherOfAIYannLecun #1BillionWithStartupBetting #ThatTheIndustryIsWrong

4

PrismLabs reports a 340% improvement in answer accuracy on multi-hop queries and a 65% reduction in hallucination rates by moving from a "naive RAG" system to a hybrid GraphRAG architecture. The main problem with the "naive RAG" was that the vector search retrieved semantically similar chunks, but wasn't able traverse the logical chain from cause to effect across document boundaries. Well, although the hybrid GraphRAG can assure causality and usually improves performance, I would challenge the numbers. 340% improvement in answer accuracy is a lot. The issue with the like-for-like comparison is that the architectures require different ways to organise information in the AI system. A system with well-organised information in the "naive RAG" architecture will outperform a system with poorly organised information in a GraphRAG system. Unfortunately, the vast majority of GraphRAG systems rely on LLMs to identify the graph elements and to build the graph. And benchmarks demonstrated that such solutions achieve even worse results than the "naive RAG" based ones. 340% improvement in answer accuracy tells a different story. Perhaps the graph was well designed, but the "naive RAG" was not. In any case, I would always strongly suggest that, if you have a thorough understanding of machine learning with graphs, go with the graphs. Otherwise, stick to simpler architectures. https://lnkd.in/dMAa3bs2

MCMC: Bridging Rendering, Optimization and Generative AI Gurprit Singh, Wenzel Jakob AMD, Max Planck Institute 2025 https://lnkd.in/gKezsqYh https://lnkd.in/gA4AeHkA In the world of AI, one of the biggest challenges is teaching machines how to *create*—to generate realistic images, scenes, or even worlds that seem to come alive. This creative process often relies on something mathematicians call sampling—drawing examples from a vast, unknown space of possibilities That’s where Markov Chain Monte Carlo (MCMC) comes in. Originally a statistical tool, MCMC has quietly become a foundational technique in both generative AI and computer graphics. It works like a patient explorer: starting from one point and taking small, probabilistic steps to map out complex, high-dimensional landscapes. This makes it ideal for helping AI systems like energy-based models (EBMs) generate authentic samples that capture the richness of the real world Interestingly, MCMC also underpins another process—gradient-based optimization, the mathematical engine that trains AI models. Each update step depends only on the current state, forming a kind of “memoryless chain” that mirrors the logic of MCMC. This conceptual overlap suggests that both optimization and sampling are two sides of the same coin Even outside AI, MCMC plays a starring role in physically based rendering, the technique that makes digital images look photorealistic by simulating how light bounces around in the real world. Sampling complex light paths is a notoriously hard problem—but MCMC’s stochastic wandering offers a solution The researchers behind this course propose that MCMC could serve as a *unifying bridge* between generative AI, optimization, and physically based rendering—fields that have traditionally evolved in parallel. By connecting them, they envision a future where machines can not only generate realistic images but also do so in ways grounded in the physical laws of light and matter For those interested, the team has made their interactive tutorials and Jupyter notebooks freely available online, inviting students and practitioners alike to explore this elegant fusion of math, physics, and imagination #MCMC #GenerativeAII #rendering #optimization #diffusionmodels #energybasedmodels #AI Exploration of how MCMC connects Generative AI, optimization, and realistic rendering 🖥️MCMC: Bridging Rendering, Optimization and Generative AI 🧠 with Dr. Gurprit Singh (AMD) and Cecile G. Tamura(Ploutos) w/ Kevin Filson 🗓️ Oct 20 · 5 PM CEST 🔗 [https://lnkd.in/gr6mpTNs) Markov Chain Monte Carlo (MCMC) isn’t just a sampling technique—it’s the quiet bridge linking how AI learns, how light behaves, and how digital worlds come to life Let's uncover how this mathematical wanderer ties together optimization, generative modeling, and physically based rendering in one unified framework

1

A new RL algorithm that fixes a hidden flaw in PPO The authors propose CE-GPPO (“Coordinating Entropy via Gradient-Preserving Policy Optimization”), a variant of PPO that restores gradient contributions from clipped actions in the policy update. They argue that traditional clipping discards useful gradient signals from low-probability tokens, which play an important role in controlling the agent’s entropy during training. By bounding those gradients in a controlled way, CE-GPPO maintains exploration–exploitation balance more stably than prior methods. They provide a theoretical analysis showing that CE-GPPO mitigates entropy instability, and empirically test it on mathematical reasoning benchmarks. Their results indicate consistent improvement over strong baselines across different model scales, demonstrating that preserving clipped gradients can lead to better performance in reinforcement learning for reasoning tasks. https://lnkd.in/gYZsJ-8f

1

📚 This paper is the third of a triptych. What began as an arithmetic substrate problem, standard autodiff imposing memory and structural costs, resolved into a composable architecture and a bootstrapping mechanism for domain-specific models from existing LLMs. arxiv.org/abs/2603.18104 The implications reach beyond training efficiency. Bayesian distillation extracts the latent prior structure of a general-purpose model and uses it to initialize domain-specific training, resolving the data-scarcity problem that makes specialized AI expensive to build. Warm rotation allows a continuously updated model to transition into an active inference pathway without service interruption, with structural correctness verified through PHG certificates. The result is a class of AI systems that are smaller, more precise, continuously adaptive, and verifiably correct to the physical structure of their respective domain. #AI #neuromorphic #posit #geometricalgebra #compilers

3

NVIDIA TensorRT LLM provides a high-level Python LLM API, with its PyTorch-native architecture enabling developers to experiment with the runtime or extend functionality. Learn how these latest TensorRT-LLM optimizations boost reasoning inference performance in our recent blog post. 🖇️ https://lnkd.in/gx52zwTA #PyTorch #OpenSourceAI #AI #Inference #Innovation

94

2 Comments