SamSam Sir

最近抽了一些時間投入短片製作，主要使用最基本的「talking photo」AI 技術，連旁白也是用 AI 模仿我的聲音來錄製。可惜目前的影片似乎偏短，接下來應該要嘗試製作長一點的內容。

這次用兩分鐘快速介紹了一個 AI Lab 的組成部分。下一次，我想實際示範如何用舊零件組裝一台 GPU 工作站——零件雖然老舊，但依然能順暢執行大型模型！

如果你也對低成本自建 AI 硬體或 AI 生成內容感興趣，歡迎收看並交流意見。🎬

建立你嘅本地AI實驗室 – 真正提升AI/數據科學事業 📌 影片簡介好多人都問：點樣先可以真係學到 AI / 數據科學實戰技能？淨係喺 cloud 度 run notebook？定係睇完理論就算？呢條片會話你知：自己起一個本地 AI 實驗室 — 即使係二手硬件，都足以令你嘅技術同事業脫胎換骨。我會.....

🚀 Building AI Factories: How Hong Kong Companies Can Create Scalable GenAI Infrastructure — On a Budget
Many Hong Kong businesses are excited about Generative AI but quickly hit a wall: high costs, complex setups, and too many promising pilots that never make it to production. Sound familiar? We call it “Pilot Purgatory.”
The good news? You don’t need massive budgets or enterprise-grade infrastructure to win with AI. The key is shifting from one-off experiments to a repeatable “AI Factory” — a modular, standardized framework for continuously building, deploying, and scaling GenAI applications.
In my latest blog post, I break down a practical blueprint tailored for Hong Kong SMEs and startups:
Leverage powerful open-source Chinese LLMs (like DeepSeek, Qwen, and Yi) that deliver excellent bilingual performance without expensive API fees.
Adopt a hybrid cloud approach — use affordable GPU instances from Tencent, Alibaba, or Huawei Cloud for training/fine-tuning, then run efficient inference on-premise or lightweight servers.
Build a clean layered architecture using tools like LangChain/LlamaIndex for orchestration, RAG with vector databases (Chroma/Weaviate), and containerized deployment with Docker + Kubernetes.
Five budget-smart strategies to escape Pilot Purgatory and turn AI into a real competitive engine.
I also share a real-world Hong Kong example: how a local sourcing company built a production system for generating product specs and contracts using a fine-tuned DeepSeek model — all for roughly HKD 30,000 in hardware plus minimal cloud costs.
Hong Kong’s unique position gives us a real edge here — strong access to high-performing Chinese models, proximity to leading cloud providers, and the agility of our business ecosystem.
If you’re a founder, CTO, or innovation leader looking to move beyond prototypes and build sustainable AI capabilities, this is for you.
👉 Read the full article here:
https://samuelsum.com/building-ai-factories-how-hong-kong-companies-can-create-scalable-genai-infrastructure-on-a-budget/
Would love to hear your thoughts — what’s the biggest barrier your company faces when scaling GenAI? Cost? Talent? Governance? Drop a comment below.
hashtag hashtag hashtag hashtag hashtag hashtag hashtag

Building AI Factories: How Hong Kong Companies Can Create Scalable GenAI Infrastructure on a Budget - Samuel Sum - Blog From Cost Centre to Competitive Engine: The "AI Factory" Mindset For Hong Kong's dynamic businesses, the promise of Generative AI (GenAI) is tempered by a harsh reality: the perceived high cost and complexity of building a robust, scalable infrastructure. Many companies find themselves trapped in "P...

AI生成「影片失敗記」，兼認真傾吓「AI Buzzword」

各位，我終於拍咗第一條廣東話片！🎉

內容係我一直好想講嘅題目：【與其追求AI神話，不如認清Buzzword濫用嘅真相】——點解而家咩都話AI？亂咁標籤對行業有咩影響？
📽️ 睇片由下面圖片連...

同場加映：本地模型生成圖片實錄
為咗身體力行「AI應用」，今次專登用本地模型生成影片配圖。結果？笑死，因為太忙無時間慢慢tune，出嚟嘅效果……都係大家自己睇啦。😅

建議大家當Podcast聽，唔好太認真睇畫面。不過對我嚟講，呢次唔太理想嘅成品本身已經係學習——至少逼自己練習咗一轉本地模型部署，算係另一種收穫！
你最近有冇試過玩AI工具但「炒車」嘅經驗？留言分享吓，等人笑吓之餘都可以一齊學嘢！😂
#廣東話內容初體驗 #數據科學

#ai生成影片失敗紀錄 #廣東話內容初體驗 #buzzword濫用 #數據科學 #ai | Samuel SUM AI生成「影片失敗記」，兼認真傾吓「AI Buzzword」 各位，我終於拍咗第一條廣東話片！🎉 內容係我一直好想講嘅題目：【與其追求AI神話，不如認清Buzzword濫用嘅真相】——點解而家咩都話AI？亂咁標籤對行業有咩影響？ 📽️ .....

由週五凌晨不停調試AI生成影片，跑到台機要大叫太熱，搞到隔離位以為火警。用把風扇仔同佢吹下風

Let's take action on how to implement AI to real-world environment -> Build your "Live" Production Cases

Agentic AI in 2026: From Hype to Real-World Deployment in Asian Enterprises - Samuel Sum - Blog Introduction: Emerging from "Pilot Purgatory" Just two years ago, in my 2024 post on "Pilot Purgatory in Machine Learning," I discussed the frustrating gap between promising prototypes and deployed production systems. Today, as we examine the state of Agentic AI in 2026, we witness a remarkable tra...

Pilot Purgatory: Why Most ML Models Shine in Prototypes But Fail in Production
As data scientists, we've all been there—building a model that crushes offline metrics in the lab, only to watch it stall forever in "pilot purgatory."
Recent reports paint a sobering picture: While classic estimates pegged 80-85% of ML projects failing to deploy, 2025 data shows persistent challenges, especially in GenAI where Gartner predicts 30% of projects will be abandoned post-POC due to costs, data quality, and unclear value. For custom GenAI tools, some studies report up to 95% failure rates in scaling beyond pilots.
The root causes? Offline-online gaps, data drift, infrastructure mismatches, skill silos, and misaligned incentives. But it's bridgeable—with a production-first mindset, early MLOps, cross-functional teams, and incremental value delivery.
I've rewritten this classic challenge from a data science lens, highlighting practical strategies to turn prototypes into scalable assets.
Read the full article here: https://lnkd.in/gTe_C4zV
What’s your biggest hurdle in deploying ML models? Share in the comments! 👇
hashtag hashtag hashtag hashtag hashtag hashtag
繁體中文版本
試點煉獄：為什麼機器學習專案在 Demo 階段表現亮眼，卻難以上線生產？
身為資料科學家，我們都經歷過——在實驗室裡建出的模型，離線指標完美無缺，卻永遠卡在「試點煉獄」無法上線。
最新報告顯示情況依然嚴峻：傳統估計有 80-85% 的 ML 專案無法部署到生產環境，而在 GenAI 領域，Gartner 預測到 2025 年底，將有 30% 的專案在 POC 階段後被放棄，原因包括成本高漲、資料品質問題及商業價值不明確。有些研究甚至指出，自訂 GenAI 工具的擴展失敗率高達 95%。
根本原因？離線與線上環境落差、資料漂移、基礎設施不匹配、技能孤島，以及激勵機制錯位。但這道鴻溝是可以跨越的——關鍵在於採用「生產優先」思維、及早導入 MLOps、跨功能團隊合作，以及漸進式價值交付。
我從資料科學專業角度重新撰寫這篇經典議題，強調實務策略，幫助將原型轉化為可擴展的資產。
完整文章請見：https://lnkd.in/gbDwZZcH
你部署 ML 模型的最大障礙是什麼？歡迎在留言區分享！👇
hashtag #機器學習 hashtag hashtag #資料科學 hashtag #人工智慧 hashtag hashtag #部署

LinkedIn This link will take you to a page that’s not on LinkedIn

It is vital to understand the value of AI agent.

AI Agents & Autonomous Systems: How AI Agents Like AutoGPT Are Evolving - Samuel Sum - Blog In the rapidly advancing world of artificial intelligence, a new frontier is emerging—AI agents and autonomous systems. These aren’t just models that respond to prompts; they are self-directed digital entities that think, plan, and act toward achieving goals with minimal human intervention.

🎧🖼️📝 Multimodal AI is redefining what it means for machines to "understand."
We're moving beyond single-mode AI systems. Today’s most advanced models—like GPT-4V, LLaVA, and Gemini—are blending text, images, audio, and even video into a single intelligence framework.
In my latest article, I explore: 🔹 What makes Multimodal AI so powerful
🔹 How GPT-4V and LLaVA work under the hood
🔹 Real-world use cases in healthcare, education, accessibility, and more
🔹 Challenges with data alignment, bias, and safety
🔹 Why Multimodal AI is a stepping stone toward AGI
As humans, we learn and interact using all our senses. Now AI can, too.
📄 Read the full article here: https://samuelsum.com/multimodal-ai-combining-text-images-and-audio-in-models-e-g-gpt-4v-llava/
Let’s talk about how this tech can make AI more helpful, accessible, and context-aware.

Multimodal AI: Combining Text, Images, and Audio in Models (e.g., GPT-4V, LLaVA) - Samuel Sum - Blog Artificial Intelligence is evolving rapidly—from processing text in chatbots to understanding images and even interpreting audio. At the forefront of this evolution is Multimodal AI: models that can process and reason across multiple data types—text, images, audio, and video—within a unified f...

Dealing with GenAI, they could help a lot but they will not do the thing for you. To take an example, I am asking the DeepSeek to convert a stored procedure... The model is giving me the "key difference" by their interpretation.

🚀 The Misconceptions About LLMs: Are Large Models Truly Omnipotent? 🤖🔍
Many companies today see large language models (LLMs) as a “magic solution” to all problems. Since the rise of DeepSeek, business leaders have rushed to invest in building localized LLM applications. But is this blind investment truly justified?
🤔 Are LLMs the right tool for every scenario?
🔹 Do companies have the necessary computing power? How do GPU limitations impact deployment?
🔹 What are the limitations of LLMs—such as computational accuracy, real-time capabilities, and hallucinations—that may hinder effectiveness?
🔹 Is an LLM alone sufficient, or should it be integrated with other AI tools, knowledge bases, and databases for better accuracy?
In my latest article, I dive into the realities of LLM adoption in enterprises and suggest more practical AI implementation strategies. Click below to read more! 👇
📖 Read the full article: https://samuelsum.com/the-misconceptions-of-llm-is-a-large-model-really-omnipotent/
中文版本(Chinese Version): https://www.linkedin.com/pulse/llm%25E7%259A%2584%25E8%25AC%258E%25E6%2580%259D%25E5%25A4%25A7%25E6%25A8%25A1%25E5%259E%258B%25E6%2598%25AF%25E5%2590%25A6%25E7%259C%259F%25E7%259A%2584%25E8%2590%25AC%25E8%2583%25BD-samuel-sum-y3yic
Would love to hear your thoughts! How should businesses take a more rational approach to LLM adoption? 💡🔗

The Misconceptions of LLM: Is a Large Model Really Omnipotent? - Samuel Sum - Blog In recent years, with the rapid development of large language models (LLMs), many corporate executives have been eagerly embracing this technology, believing it to be a panacea for all problems. Since early 2025, the rise of DeepSeek has further fueled market enthusiasm, especially in Hong Kong, whe...