<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Billy Tse Blog</title>
    <link>https://billytse.dev/blog</link>
    <description>Technical articles about AI, machine learning, quantum computing, and technology by Billy Tse.</description>
    <language>en-US</language>
    <lastBuildDate>Mon, 24 Aug 2026 07:00:39 GMT</lastBuildDate>
    <atom:link href="https://billytse.dev/feed.xml" rel="self" type="application/rss+xml"/>
    <image>
      <url>https://billytse.dev/og-image.jpg</url>
      <title>Billy Tse Blog</title>
      <link>https://billytse.dev/blog</link>
    </image>
    
    <item>
      <title><![CDATA[YC Fall 2026 RFS：13 個創業趨勢分析]]></title>
      <link>https://billytse.dev/blog/yc-fall-2026-rfs-startup-trends-analysis</link>
      <guid isPermaLink="true">https://billytse.dev/blog/yc-fall-2026-rfs-startup-trends-analysis</guid>
      <description><![CDATA[拆解 Y Combinator Fall 2026 Requests for Startups：由 AI tutor、multiplayer agents、small software，到 defense、海上 data center、human verification，分析 13 個方向背後共同嘅技術、成本、人口、監管同地緣政治趨勢。]]></description>
      <pubDate>Tue, 04 Aug 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>Hardware</category>
      <category>Privacy</category>
      <category>Web Development</category>
    </item>
    <item>
      <title><![CDATA[AI Agent Skills、MCP 同 Rules：三者有咩分別？]]></title>
      <link>https://billytse.dev/blog/ai-agent-skills-mcp-rules-difference</link>
      <guid isPermaLink="true">https://billytse.dev/blog/ai-agent-skills-mcp-rules-difference</guid>
      <description><![CDATA[用餐廳守則、食譜同供應商連線做例子，簡單分清 AI Agent Skills、MCP 與 Rules 各自負責咩。]]></description>
      <pubDate>Tue, 04 Aug 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>DevOps</category>
    </item>
    <item>
      <title><![CDATA[DeepSeek Reasonix：以 Prefix Cache 為核心嘅 AI Coding Agent]]></title>
      <link>https://billytse.dev/blog/deepseek-reasonix-prefix-cache-ai-coding-agent</link>
      <guid isPermaLink="true">https://billytse.dev/blog/deepseek-reasonix-prefix-cache-ai-coding-agent</guid>
      <description><![CDATA[用茶餐廳落單同書枱整理等生活例子，簡單理解 Reasonix 點樣靠 prefix cache 降低 AI coding 長對話成本。]]></description>
      <pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>DevOps</category>
    </item>
    <item>
      <title><![CDATA[DSpark：DeepSeek 點樣「一次估多個、驗證前綴、錯咗即修正」令 V4 快 60–85%？]]></title>
      <link>https://billytse.dev/blog/deepseek-dspark-confidence-scheduled-speculative-decoding</link>
      <guid isPermaLink="true">https://billytse.dev/blog/deepseek-dspark-confidence-scheduled-speculative-decoding</guid>
      <description><![CDATA[DeepSeek × 北大推出 DSpark：用 parallel backbone 一次估多個 tokens，再用輕量 Markov head 修正前後依賴，配合 confidence scheduler 按 GPU 負載決定驗證幾長。本文拆解它如何在保持 target distribution 不變下，令 DeepSeek-V4 每位用戶生成速度提升 57–85%。]]></description>
      <pubDate>Thu, 30 Jul 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>NLP</category>
      <category>Inference Optimization</category>
      <category>Hardware Acceleration</category>
    </item>
    <item>
      <title><![CDATA[Kalman Filter 深度拆解：完整指南]]></title>
      <link>https://billytse.dev/blog/kalman-filter-deep-dive-from-apollo-to-sort</link>
      <guid isPermaLink="true">https://billytse.dev/blog/kalman-filter-deep-dive-from-apollo-to-sort</guid>
      <description><![CDATA[由 1960 年 Rudolf E. Kálmán 嘅原始 paper 講起，徹底拆解 Kalman Filter 嘅 Bayesian 推導、Predict-Update cycle、Kalman Gain 嘅幾何意義，再延伸到 EKF / UKF、SORT/DeepSORT 入面嘅 constant-velocity model、SLAM 同 IMU sensor fusion 嘅實戰應用。附 NumPy + PyTorch 實作。]]></description>
      <pubDate>Mon, 25 May 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>Computer Vision</category>
    </item>
    <item>
      <title><![CDATA[RTMO：完整指南]]></title>
      <link>https://billytse.dev/blog/rtmo-one-stage-realtime-multi-person-pose-estimation</link>
      <guid isPermaLink="true">https://billytse.dev/blog/rtmo-one-stage-realtime-multi-person-pose-estimation</guid>
      <description><![CDATA[深入拆解 RTMO（CVPR 2024）：第一個將 dual 1-D heatmap coordinate classification 整合落 YOLO 嘅 one-stage real-time multi-person pose estimation framework。Dynamic Coordinate Classifier（DBA + DBE）、MLE loss with learnable variance、YOLOX-style training，喺 COCO val2017 攞到 74.8% AP + 141 FPS（V100）。]]></description>
      <pubDate>Sun, 24 May 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Computer Vision</category>
      <category>AI</category>
      <category>Image Processing</category>
    </item>
    <item>
      <title><![CDATA[DFlash × DDTree：Block Diffusion 推測解碼解析]]></title>
      <link>https://billytse.dev/blog/dflash-ddtree-luce-dflash-block-diffusion-speculative-decoding-2026</link>
      <guid isPermaLink="true">https://billytse.dev/blog/dflash-ddtree-luce-dflash-block-diffusion-speculative-decoding-2026</guid>
      <description><![CDATA[2026 年 2 月 z-lab 出咗 DFlash（block diffusion drafter）6× lossless speculative decoding；4 月 Liran Ringel 同 Yaniv Romano 出咗 DDTree，在 DFlash 上面加咗 tree-structured verify，全面 60/60 setting 都提升；Luce-Org 跨咗 GGUF + ggml + 3 個 custom CUDA kernel，俾一部 24 GB RTX 3090 跑 Qwen3.5-27B 跑到 207 tok/s peak、3.43× 加速。今次拆解三個層次：paper 原理、tree verification math、同 consumer GPU 部署實戰。]]></description>
      <pubDate>Wed, 29 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>NLP</category>
      <category>Inference Optimization</category>
      <category>Hardware Acceleration</category>
      <category>GPU</category>
      <category>Diffusion</category>
    </item>
    <item>
      <title><![CDATA[LLM Wiki vs RAG：Karpathy 知識編譯範式解析]]></title>
      <link>https://billytse.dev/blog/llm-wiki-vs-rag-karpathy-knowledge-compilation-2026</link>
      <guid isPermaLink="true">https://billytse.dev/blog/llm-wiki-vs-rag-karpathy-knowledge-compilation-2026</guid>
      <description><![CDATA[Andrej Karpathy 喺 2026 年 4 月放咗一份 GitHub Gist 講 LLM Wiki：用一個 raw/、wiki/、index.md 嘅 Markdown 結構，配合 LLM 自動編譯同維護，挑戰 traditional RAG 嘅 chunk + embed + retrieve pipeline。今次深入拆解 compile-time vs query-time 嘅範式分別、stateful 知識點樣 compound、80 年前 Memex 點解失敗、同點解 LLM Wiki 唔係 RAG 嘅替代品而係另一隻完全唔同嘅嘢。]]></description>
      <pubDate>Wed, 29 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>NLP</category>
    </item>
    <item>
      <title><![CDATA[MemPalace：完整指南]]></title>
      <link>https://billytse.dev/blog/mempalace-method-of-loci-llm-long-term-memory-chromadb-2026</link>
      <guid isPermaLink="true">https://billytse.dev/blog/mempalace-method-of-loci-llm-long-term-memory-chromadb-2026</guid>
      <description><![CDATA[MemPalace 喺 2026 年 4 月被 Milla Jovovich 同 Ben Sigman 開源，14 日內爆 47k GitHub stars。佢將古希臘嘅 method of loci（記憶宮殿）搬入 LLM agent 嘅長期記憶系統，配 ChromaDB + SQLite + 29 個 MCP tools，喺 LongMemEval 攞到 96.6% R@5（hybrid mode 100%），全部 zero API cost、純本地運行。今篇深入拆 Wings/Rooms/Closets/Drawers 嘅 4-layer 架構、AAAK compression、PALACE_PROTOCOL prompt 設計、Knowledge Graph，再客觀分析點解 raw mode 跑得仲快過 palace structure，同其他 memory 系統（Mem0、Zep、Supermemory）點比。]]></description>
      <pubDate>Tue, 28 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>NLP</category>
      <category>Inference Optimization</category>
    </item>
    <item>
      <title><![CDATA[Coarse Solver × Fine Solver：完整指南]]></title>
      <link>https://billytse.dev/blog/parareal-coarse-fine-solver-neural-parareal-randnet-2026</link>
      <guid isPermaLink="true">https://billytse.dev/blog/parareal-coarse-fine-solver-neural-parareal-randnet-2026</guid>
      <description><![CDATA[並行運算入面，空間並行係常識，但時間呢條軸點 parallelize？Parareal algorithm（Lions, Maday, Turinici 2001）用一個 cheap 嘅 coarse solver G + 一個 expensive 嘅 fine solver F，配合 predictor-corrector iteration，將 sequential time integration 變成 parallel-in-time。深入拆解 coarse / fine solver 嘅角色、convergence、speedup bound，再睇 PPINN（Karniadakis 2020）、Neural-Parareal（ITER 2024）、RandNet-Parareal（NeurIPS 2024）等用 neural network 取代 G 嘅最新發展。]]></description>
      <pubDate>Mon, 27 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>Inference Optimization</category>
    </item>
    <item>
      <title><![CDATA[DeepSeek-V4 架構解析：1M Context 與 Hybrid Attention]]></title>
      <link>https://billytse.dev/blog/deepseek-v4-hybrid-attention-csa-hca-mhc-muon-1m-context-2026</link>
      <guid isPermaLink="true">https://billytse.dev/blog/deepseek-v4-hybrid-attention-csa-hca-mhc-muon-1m-context-2026</guid>
      <description><![CDATA[DeepSeek 喺 2026 年 4 月 24 日 release 咗 V4 系列（V4-Pro 1.6T / 49B active；V4-Flash 284B / 13B active），全部支援 1M token context。深入拆解 5 個核心架構創新：CSA（4× 壓縮 + lightning indexer 揀 top-1024）、HCA（128× 壓縮 + dense attention）、manifold-constrained hyper-connections（mHC，用 Birkhoff Polytope + Sinkhorn-Knopp 取代 residual）、Muon optimizer、FP4 QAT，加埋 agent-side 嘅 interleaved thinking、|DSML| XML tool calls、DSec RL sandbox。]]></description>
      <pubDate>Sat, 25 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>Transformer</category>
      <category>NLP</category>
      <category>Inference Optimization</category>
      <category>Attention Mechanisms</category>
    </item>
    <item>
      <title><![CDATA[Recurrent-Depth Transformers：Parcae 與 LTG 完整解析]]></title>
      <link>https://billytse.dev/blog/parcae-loop-think-generalize-recurrent-depth-transformers-2026</link>
      <guid isPermaLink="true">https://billytse.dev/blog/parcae-loop-think-generalize-recurrent-depth-transformers-2026</guid>
      <description><![CDATA[深入解析 2026 年 4 月呢個星期同時登場嘅兩篇 Recurrent-Depth Transformers 論文——Parcae（UCSD + Together AI，用 LTI 系統 + 負對角參數化徹底解決 looped model 訓練崩潰，首個 RDT scaling laws）同 Loop, Think, & Generalize（OSU NLP，首次系統驗證 RDT 做到 implicit 多跳推理同 depth extrapolation）。附八年進化背景、完整 PyTorch 骨架、實作心得。]]></description>
      <pubDate>Wed, 22 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>Transformer</category>
      <category>NLP</category>
      <category>Attention Mechanisms</category>
    </item>
    <item>
      <title><![CDATA[由 Prompt Engineering 到 Context Engineering 到 Harness Engin]]></title>
      <link>https://billytse.dev/blog/prompt-context-harness-engineering-evolution</link>
      <guid isPermaLink="true">https://billytse.dev/blog/prompt-context-harness-engineering-evolution</guid>
      <description><![CDATA[深入解析 AI 工程由 Prompt Engineering（2022-24）到 Context Engineering（2025）到 Harness Engineering（2026）嘅進化歷程。了解 Karpathy、Tobi Lütke、Anthropic、OpenAI、LangChain 點樣定義呢三個時代，附完整實戰指南同代碼範例。]]></description>
      <pubDate>Mon, 06 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>NLP</category>
    </item>
    <item>
      <title><![CDATA[Person ReID 進化史：完整指南]]></title>
      <link>https://billytse.dev/blog/person-reid-evolution-transreid-solider-dinov2</link>
      <guid isPermaLink="true">https://billytse.dev/blog/person-reid-evolution-transreid-solider-dinov2</guid>
      <description><![CDATA[深入解析 Person ReID 由 CNN 時代到 Transformer 時代嘅三大里程碑：TransReID（首個 pure Transformer ReID）、SOLIDER（語義可控自監督預訓練）、DINOv2（通用視覺基礎模型）。了解 JPM、SIE、語義控制器等核心設計，附完整實作指南。]]></description>
      <pubDate>Sun, 05 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Computer Vision</category>
      <category>AI</category>
      <category>Transformer</category>
      <category>Image Processing</category>
    </item>
    <item>
      <title><![CDATA[OSNet：點解一個 2.2M 參數嘅輕量 CNN 可以打贏 ResNet50 做 Person Re-ID？]]></title>
      <link>https://billytse.dev/blog/osnet-omni-scale-feature-learning-person-reid</link>
      <guid isPermaLink="true">https://billytse.dev/blog/osnet-omni-scale-feature-learning-person-reid</guid>
      <description><![CDATA[深入解析 OSNet（Omni-Scale Network），一個專為行人重識別設計嘅輕量 CNN。了解點樣透過 multi-scale streams + unified aggregation gate 學習 omni-scale features，用 2.2M 參數打贏 24M 嘅 ResNet50。]]></description>
      <pubDate>Sat, 04 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Computer Vision</category>
      <category>AI</category>
      <category>Image Processing</category>
    </item>
    <item>
      <title><![CDATA[乜嘢係 1×1 Convolution？點解佢可以又慳參數又做跨 Channel Mix？]]></title>
      <link>https://billytse.dev/blog/1x1-convolution</link>
      <guid isPermaLink="true">https://billytse.dev/blog/1x1-convolution</guid>
      <description><![CDATA[用廣東話拆解 1×1 convolution（pointwise conv）：由 Network in Network 到 ResNet bottleneck，再到 MobileNet depthwise separable，講清楚佢點樣做 channel mixing、降/升維同加速。]]></description>
      <pubDate>Wed, 01 Apr 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Computer Vision</category>
      <category>AI</category>
    </item>
    <item>
      <title><![CDATA[TurboQuant：Google 用 3-bit 壓縮 KV Cache 做到零精度損失嘅突破]]></title>
      <link>https://billytse.dev/blog/turboquant-online-vector-quantization-near-optimal</link>
      <guid isPermaLink="true">https://billytse.dev/blog/turboquant-online-vector-quantization-near-optimal</guid>
      <description><![CDATA[深入解析 Google Research 發表嘅 TurboQuant 論文，了解點樣透過 PolarQuant + QJL 兩階段量化，將 LLM KV cache 壓縮到 3-bit 而保持零精度損失，仲有 8x speedup]]></description>
      <pubDate>Thu, 26 Mar 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Inference Optimization</category>
      <category>AI</category>
      <category>Transformer</category>
    </item>
    <item>
      <title><![CDATA[Quantum Teleportation：瞬間移動嘅數學與物理（連埋 Cheat Sheet 拆解）]]></title>
      <link>https://billytse.dev/blog/quantum-teleportation-math-physics-cheat-sheet</link>
      <guid isPermaLink="true">https://billytse.dev/blog/quantum-teleportation-math-physics-cheat-sheet</guid>
      <description><![CDATA[完整數學推導 Quantum Teleportation（量子隱形傳態）嘅原理，從 Bell State 建立到測量修正，附帶實用 Cheat Sheet 同埋直觀比喻解釋。]]></description>
      <pubDate>Mon, 09 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Quantum</category>
      <category>CMSC 5711</category>
      <category>AI</category>
    </item>
    <item>
      <title><![CDATA[BullMQ vs Redis Streams vs RabbitMQ：完整比較]]></title>
      <link>https://billytse.dev/blog/bullmq-redis-rabbitmq-nodejs-job-queue-comparison</link>
      <guid isPermaLink="true">https://billytse.dev/blog/bullmq-redis-rabbitmq-nodejs-job-queue-comparison</guid>
      <description><![CDATA[深入對比 Redis Pub/Sub、Redis Streams、BullMQ 同 RabbitMQ 四大 Message Queue 方案，解析點解 Redis Pub/Sub 唔適合做 Job Queue，Redis Streams 點樣處理持久化消息流，BullMQ 點樣構建可靠嘅背景任務處理系統，RabbitMQ 幾時先係最佳選擇，附完整實戰代碼同場景推薦]]></description>
      <pubDate>Fri, 06 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>NLP</category>
      <category>Web Development</category>
    </item>
    <item>
      <title><![CDATA[LLM Pre-training 架構詳解：原理與應用]]></title>
      <link>https://billytse.dev/blog/llm-pretraining-architectures-decoder-encoder</link>
      <guid isPermaLink="true">https://billytse.dev/blog/llm-pretraining-architectures-decoder-encoder</guid>
      <description><![CDATA[深入探討 LLM pre-training 嘅三大架構：Decoder-Only (GPT)、Encoder-Only (BERT) 同 Encoder-Decoder (T5)，分析佢哋嘅訓練方式、優缺點、實際應用場景，同埋點解 Decoder-Only 成為主流。]]></description>
      <pubDate>Thu, 05 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>AI</category>
      <category>NLP</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[Telegram vs WhatsApp：點解 Telegram 咁注重私隱？]]></title>
      <link>https://billytse.dev/blog/telegram-vs-whatsapp-privacy-comparison</link>
      <guid isPermaLink="true">https://billytse.dev/blog/telegram-vs-whatsapp-privacy-comparison</guid>
      <description><![CDATA[深入比較 Telegram 同 WhatsApp 嘅私隱保護機制，了解點解 Telegram 被視為更私隱導向嘅通訊平台，分析加密方式、數據收集、同埋伺服器架構嘅分別。]]></description>
      <pubDate>Thu, 05 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Privacy</category>
      <category>Personal</category>
    </item>
    <item>
      <title><![CDATA[FlashAttention 1, 2, 3, 4 完全解析：完整指南]]></title>
      <link>https://billytse.dev/blog/flashattention-1-2-3-complete-guide</link>
      <guid isPermaLink="true">https://billytse.dev/blog/flashattention-1-2-3-complete-guide</guid>
      <description><![CDATA[深入解析 FlashAttention 四代演進：從 FlashAttention 1 嘅 IO-aware tiling，到 FlashAttention 2 嘅並行優化，再到 FlashAttention 3 嘅異步計算同 FP8 支援，最後到 FlashAttention 4 喺 Blackwell B200 突破 1.6 PetaFLOPs，了解點樣將 Transformer Attention 推向極致]]></description>
      <pubDate>Wed, 04 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Attention Mechanisms</category>
      <category>Inference Optimization</category>
      <category>Hardware Acceleration</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[解構 WanVideo 2.2：雙專家接力嘅 AI 影片生成工作流]]></title>
      <link>https://billytse.dev/blog/wanvideo-2-2-moe-workflow</link>
      <guid isPermaLink="true">https://billytse.dev/blog/wanvideo-2-2-moe-workflow</guid>
      <description><![CDATA[深入解析 WanVideo 2.2 點樣透過 Mixture-of-Experts (MoE) 架構，用高噪點同低噪點兩個專家模型接力，喺 ComfyUI 生成電影級 AI 影片]]></description>
      <pubDate>Wed, 04 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Video Generation</category>
      <category>Diffusion</category>
      <category>Image Processing</category>
    </item>
    <item>
      <title><![CDATA[Screen 命令完全指南：點樣喺 SSH 斷線後繼續跑緊你嘅程式]]></title>
      <link>https://billytse.dev/blog/linux-screen-command-guide</link>
      <guid isPermaLink="true">https://billytse.dev/blog/linux-screen-command-guide</guid>
      <description><![CDATA[深入淺出講解 Linux Screen command，包括基本用法、實戰場景、Screen vs tmux 對比，以及生產環境最佳實踐]]></description>
      <pubDate>Wed, 04 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Linux</category>
      <category>DevOps</category>
    </item>
    <item>
      <title><![CDATA[Glyph 同 Vist：點解將文字變圖像反而更慳 Tokens？]]></title>
      <link>https://billytse.dev/blog/glyph-vist-text-to-image-token-compression</link>
      <guid isPermaLink="true">https://billytse.dev/blog/glyph-vist-text-to-image-token-compression</guid>
      <description><![CDATA[深入探討 Glyph 同 Vist 兩篇突破性論文，了解點樣透過將文字 render 成圖像再用 vision encoder 壓縮，實現 slow-fast 混合架構處理超長語境]]></description>
      <pubDate>Tue, 03 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Visual Tokens</category>
      <category>OCR</category>
      <category>Transformer</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[AI 工作站選購：RTX 6000 Blackwell 與 Threadripper]]></title>
      <link>https://billytse.dev/blog/ai-workstation-rtx-6000-blackwell-threadripper-cooling</link>
      <guid isPermaLink="true">https://billytse.dev/blog/ai-workstation-rtx-6000-blackwell-threadripper-cooling</guid>
      <description><![CDATA[深入探討 AI 訓練工作站嘅硬件配置，包括 NVIDIA RTX 6000 Pro Blackwell、AMD Threadripper、PCIe 5.0、多 GPU 配置，以及雙風扇、渦輪式、被動散熱等冷卻方案嘅技術細節]]></description>
      <pubDate>Tue, 03 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>GPU</category>
      <category>Hardware</category>
    </item>
    <item>
      <title><![CDATA[DeepSeek OCR：完整指南]]></title>
      <link>https://billytse.dev/blog/deepseek-ocr-visual-token-efficiency</link>
      <guid isPermaLink="true">https://billytse.dev/blog/deepseek-ocr-visual-token-efficiency</guid>
      <description><![CDATA[Discover how DeepSeek OCR achieves state-of-the-art document parsing with only 1120 visual tokens—6x fewer than competitors—through innovative visual causal flow and semantic reordering.]]></description>
      <pubDate>Tue, 03 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>OCR</category>
      <category>Visual Tokens</category>
      <category>Transformer</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[量子電路入門：Dirac Notation、量子閘同 Hadamard Gate]]></title>
      <link>https://billytse.dev/blog/quantum-circuits-intro-gates-hadamard</link>
      <guid isPermaLink="true">https://billytse.dev/blog/quantum-circuits-intro-gates-hadamard</guid>
      <description><![CDATA[深入探討量子電路嘅基礎概念，包括 Dirac notation（bra-ket）、常見量子閘（Pauli Gates、CNOT、SWAP）、Hadamard Gate 同埋點樣用量子電路構建量子演算法]]></description>
      <pubDate>Mon, 02 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Quantum</category>
    </item>
    <item>
      <title><![CDATA[NVIDIA DGX Spark + Mac Studio：EXO 1.0 混合架構實現 4 倍 LLM 加速]]></title>
      <link>https://billytse.dev/blog/nvidia-dgx-spark-mac-studio-exo</link>
      <guid isPermaLink="true">https://billytse.dev/blog/nvidia-dgx-spark-mac-studio-exo</guid>
      <description><![CDATA[深入拆解 EXO Labs 最新方案：點樣結合 NVIDIA DGX Spark 嘅強大算力同 Mac Studio 嘅高記憶體頻寬，透過分離 Prefill 同 Decode 階段，實現 4 倍 LLM 推理速度提升。]]></description>
      <pubDate>Sun, 01 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Inference Optimization</category>
      <category>GPU</category>
      <category>Hardware</category>
    </item>
    <item>
      <title><![CDATA[Lexapro (Escitalopram) 完整指南：SSRI、MDD 同用藥須知]]></title>
      <link>https://billytse.dev/blog/lexapro-ssri-mdd-guide</link>
      <guid isPermaLink="true">https://billytse.dev/blog/lexapro-ssri-mdd-guide</guid>
      <description><![CDATA[深入了解 Lexapro (Escitalopram) 嘅作用機制、SSRI 藥理、重性抑鬱症 (MDD) 成因同類型、藥物相互作用、安全用藥指引等全面資訊。]]></description>
      <pubDate>Sun, 01 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Mental Health</category>
    </item>
    <item>
      <title><![CDATA[Next.js Server-Side Rendering 同 SEO：完整指南]]></title>
      <link>https://billytse.dev/blog/nextjs-ssr-seo-force-dynamic</link>
      <guid isPermaLink="true">https://billytse.dev/blog/nextjs-ssr-seo-force-dynamic</guid>
      <description><![CDATA[深入探討 Next.js 三種 rendering 策略（SSG、SSR、CSR）對 SEO 嘅影響，解析 force-dynamic 點樣喺 server-side 生成完整 HTML，確保搜尋引擎可以正確 crawl 你嘅內容]]></description>
      <pubDate>Sun, 01 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Next.js</category>
      <category>SEO</category>
      <category>Web Development</category>
    </item>
    <item>
      <title><![CDATA[State Space Models (SSM)：由 Mamba 到 S4 嘅演變]]></title>
      <link>https://billytse.dev/blog/state-space-models-explained</link>
      <guid isPermaLink="true">https://billytse.dev/blog/state-space-models-explained</guid>
      <description><![CDATA[深入探討 State Space Models 的原理、架構演變，以及點解佢哋可以挑戰 Transformer 的地位]]></description>
      <pubDate>Sun, 01 Feb 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>SSM / Mamba</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[DeepSeek-OCR 2：點樣教 AI 好似人咁「睇」文件？Visual Causal Flow 完全拆解]]></title>
      <link>https://billytse.dev/blog/deepseek-ocr-2-visual-causal-flow</link>
      <guid isPermaLink="true">https://billytse.dev/blog/deepseek-ocr-2-visual-causal-flow</guid>
      <description><![CDATA[用廣東話深入拆解 DeepSeek-OCR 2 論文：由 raster-scan 嘅根本問題、DeepEncoder V2 嘅 dual-stream attention、causal flow queries 到 cascade causal reasoning，附具體數字例子、attention mask 推導、同完整 inference code。]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>OCR</category>
      <category>Visual Tokens</category>
      <category>Transformer</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[Recursive Language Models (RLM)：MIT 突破超長文檔處理嘅新架構]]></title>
      <link>https://billytse.dev/blog/recursive-language-models-rlm-mit</link>
      <guid isPermaLink="true">https://billytse.dev/blog/recursive-language-models-rlm-mit</guid>
      <description><![CDATA[深入探討 MIT 最新推出的 Recursive Language Models (RLM) 架構，了解點樣透過遞歸調用解決 context rot 問題，處理 1000 萬+ tokens 嘅超長文檔]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Transformer</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[Attention 機制進化史：由 2014 到 2026 嘅演變]]></title>
      <link>https://billytse.dev/blog/attention-mechanisms-timeline</link>
      <guid isPermaLink="true">https://billytse.dev/blog/attention-mechanisms-timeline</guid>
      <description><![CDATA[全面回顧 Attention 機制從誕生到現在的演化歷程，分析哪些方法仍在使用，哪些已經被淘汰]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Attention Mechanisms</category>
      <category>Transformer</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[Transformer 架構詳解：Attention、QKV 同 Multi-Head 機制]]></title>
      <link>https://billytse.dev/blog/transformer-architecture-explained</link>
      <guid isPermaLink="true">https://billytse.dev/blog/transformer-architecture-explained</guid>
      <description><![CDATA[深入淺出講解 Transformer 架構，包括 Self-Attention 機制、Query/Key/Value 概念、Encoder-Decoder 設計，以及 Multi-Head Attention 嘅運作原理]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Transformer</category>
      <category>Attention Mechanisms</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[Rotary Position Embeddings (RoPE)]]></title>
      <link>https://billytse.dev/blog/rotary-position-embeddings</link>
      <guid isPermaLink="true">https://billytse.dev/blog/rotary-position-embeddings</guid>
      <description><![CDATA[深入探討 Rotary Position Embeddings (RoPE) 的原理、數學推導及其在 Transformer 模型中的應用]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Position Encoding</category>
      <category>Transformer</category>
      <category>CSCI 5640 NLP</category>
    </item>
    <item>
      <title><![CDATA[霍夫變換 (Hough Transform)：直線與圓形檢測]]></title>
      <link>https://billytse.dev/blog/hough-transform</link>
      <guid isPermaLink="true">https://billytse.dev/blog/hough-transform</guid>
      <description><![CDATA[深入探討 Hough Transform 的原理、數學推導及實際應用，包括直線檢測、圓形檢測與 Linear Regression 的比較]]></description>
      <pubDate>Wed, 28 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Computer Vision</category>
      <category>Image Processing</category>
      <category>CMSC 5711</category>
    </item>
    <item>
      <title><![CDATA[量子退火 (Quantum Annealing)]]></title>
      <link>https://billytse.dev/blog/quantum-annealing</link>
      <guid isPermaLink="true">https://billytse.dev/blog/quantum-annealing</guid>
      <description><![CDATA[量子退火核心原理（如能量地景與量子穿隧效應）、應用領域以及參考來源]]></description>
      <pubDate>Mon, 26 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Quantum</category>
    </item>
    <item>
      <title><![CDATA[量子計算：未來科技的新紀元]]></title>
      <link>https://billytse.dev/blog/quantum-computing-future</link>
      <guid isPermaLink="true">https://billytse.dev/blog/quantum-computing-future</guid>
      <description><![CDATA[深入探討量子計算的原理、應用及其對未來科技發展的影響]]></description>
      <pubDate>Mon, 26 Jan 2026 00:00:00 GMT</pubDate>
      <author>Billy Tse</author>
      <category>Quantum</category>
    </item>
  </channel>
</rss>