https://subhash.net/tags/machine-learning/Recent content in Machine Learning on Subhash DasyamHugoenTue, 06 Jan 2026 12:47:27 +0400https://subhash.net/posts/decoding-fp32-fp16-fp8-int8-int4-master/Thu, 14 Aug 2025 00:41:00 +0400https://subhash.net/posts/decoding-fp32-fp16-fp8-int8-int4-master/<h2 id="the-master-chefs-dilemma-understanding-precision-in-a-world-of-efficiency">The Master Chef&rsquo;s Dilemma: Understanding Precision in a World of Efficiency</h2> <h2 id="every-executives-nightmare">Every Executive&rsquo;s Nightmare</h2> <p>Picture this: You&rsquo;re running the world&rsquo;s most exclusive restaurant chain. Your head chef is a genius - creates absolutely perfect dishes every single time. But there&rsquo;s a catastrophic problem that&rsquo;s bleeding your company dry.</p>https://subhash.net/posts/mixture-of-experts-moe-specialist/Sun, 13 Jul 2025 23:00:00 +0400https://subhash.net/posts/mixture-of-experts-moe-specialist/<p>Building on our transformer story - if you haven&rsquo;t read the complete transformer guide yet, check it out first!</p> <h2 id="remember-our-transformer-story">Remember Our Transformer Story?</h2> <p>In our previous deep dive, we learned that transformers have this amazing &ldquo;deep thinking step&rdquo; (the Feed Forward Network) where they:</p>https://subhash.net/posts/how-transformers-actually-work-complete/Tue, 08 Jul 2025 00:45:00 +0400https://subhash.net/posts/how-transformers-actually-work-complete/<p>Ever wondered how ChatGPT, Claude, or GPT-4 actually understand and generate text? Let me break down the magic behind transformers like you&rsquo;re 12 years old! 👇</p> <p>Note: When I mention &ldquo;117 million parameters&rdquo; in examples, I&rsquo;m talking about GPT-1 and BERT-base models. Modern models like GPT-4 are much, much bigger!</p>https://subhash.net/posts/rag-revolution-how-application-aware/Tue, 17 Jun 2025 14:48:00 +0400https://subhash.net/posts/rag-revolution-how-application-aware/<h2 id="paper-review-and-attribution">Paper Review and Attribution</h2> <p>This article is based on the fascinating research paper &ldquo;RAG+: Enhancing Retrieval-Augmented Generation with Application-Aware Reasoning&rdquo; by Yu Wang, Shiwan Zhao, Ming Fan, and colleagues from Huawei Technologies, Xi&rsquo;an Jiaotong University, and Nankai University.</p>https://subhash.net/posts/graceful-degradation-strategies-for/Sun, 15 Jun 2025 12:40:00 +0400https://subhash.net/posts/graceful-degradation-strategies-for/<h2 id="introduction">Introduction</h2> <p>Graceful degradation ensures systems maintain core functionality even when components fail or face performance issues, rather than experiencing complete system failure. In GenAI and inference systems, this capability becomes mission-critical as organizations increasingly rely on AI-powered applications for business operations. The approach involves systematically reducing less critical services while preserving essential operations during high-stress conditions or failures.</p>https://subhash.net/posts/agentic-ai-using-buzzword-to-justify/Fri, 06 Jun 2025 13:00:00 +0400https://subhash.net/posts/agentic-ai-using-buzzword-to-justify/<p>Since my original post took off, quite a few of you have reached out asking for more detailed examples. So today, I’m diving into one of those examples from the previous post and unpacking it in greater depth.</p>https://subhash.net/posts/the-complete-guide-to-transformer/Wed, 21 May 2025 22:58:00 +0400https://subhash.net/posts/the-complete-guide-to-transformer/<h3 id="1-the-big-breakthrough-introduction-to-transformer-architecture">1. The Big Breakthrough: Introduction to Transformer Architecture</h3> <p>Two weeks after successfully implementing his first transformer model, Alex was hunched over his laptop in the university AI lab, a look of amazement on his face as he compared the results from his old model and his new transformer-based solution.</p>