Dr.Jiw

Youtube, FB, IG contents are more dynamic because of their varying URLs caused by CDNs. The URLs are the key used by Squid to identify cached objects. IMHO, an optimal solution is to hash the contents to let Cache search by content hash values rather than content URLs. However, how to know the hash value requested by users in advance?

BTW, "the YouTube URL can be parsed to find the id of the video. These seem to be constant for the same video of the same resolution. "

I've tried on this observation using Laptop and Smartphone to open the same YouTube clip:

Laptop: https://www.youtube.com/watch?v=oAwGBYdvwW0
Smartphone: https://www.youtube.com/oAwGBYdvwW0?si=qmdgeTLhUYobTXi6
We can see the video ID "oAwGBYdvwW0"

Facebook to the same image:

Laptop: https://scontent.fbkk6-2.fna.fbcdn.net/v/t39.30808-6/475548702_1142874757632916_6715193667665471540_n.jpg?_nc_cat=109&ccb=1-7&_nc_sid=127cfc&_nc_ohc=97bVWxJsE34Q7kNvgGXLIZG&_nc_zt=23&_nc_ht=scontent.fbkk6-2.fna&_nc_gid=A4HSUZOcURM3gh5kv_J0jUd&oh=00_AYCL-mz6AA3IVSH3ZZfKOFrqcJeT8NqjvSgC6VJYPJ8R5g&oe=67A756B0
Smartphone (desktop mode): https://scontent.fbkk6-2.fna.fbcdn.net/v/t39.30808-6/475548702_1142874757632916_6715193667665471540_n.jpg?_nc_cat=109&ccb=1-7&_nc_sid=127cfc&_nc_ohc=97bVWxJsE34Q7kNvgGXLIZG&_nc_zt=23&_nc_ht=scontent.fbkk6-2.fna&_nc_gid=AxT6EmOD0b2GQ9OOKogzORK&oh=00_AYDAk_MK9VhJ4oyrHOLXz8HOmAmo4AeoHfjqxGjDW7rU8Q&oe=67A756B0
Image ID is probably "475548702_1142874757632916_6715193667665471540_n.jpg" ซึ่งโชว์อยู๋บนชื่อ tab ใน web browser ด้วย

เพราะยังไงก็ตาม dynamic URLs ต้อง map ไปยัง the same resource object (many to one) อยู่แล้วโดยอาศัย Object ID นั่งเอง จึงต้อง parse object ID ออกมาให้ได้ก็จะ cache ได้ในที่สุด ส่วน one to many หมายถึง URL เดียวอาจ refer to different objects แต่ต้อง ณ ต่างเวลากันซึ่งเกิดจาก URL reuse ซึ่งก็เข้าข่าย object expiration ของ object เดิมก่อนที่ URL จะถูก reused ไปชี้ new object

ท้ายที่สุดเป็นเรื่องของการ implement Caching proxy ทำยังไงจึง parse Object ID to enable caching ส่วน cache replacement policy ยังมีประโยชน์เหมือนเดิม

https://wiki.squid-cache.org/ConfigExamples/DynamicContent/YouTube

https://wiki.squid-cache.org/ConfigExamples/DynamicContent/Coordinator

วันเสาร์ที่ 1 กุมภาพันธ์ พ.ศ. 2568

Tractable complexity classes

In computational complexity theory, the complexity class "P" (Polynomial time) is considered the class of tractable problems, meaning problems that can be solved by an algorithm with a running time that is polynomial in the size of the input, making them considered "efficiently solvable" in practice. --Gemini

List of complexity classes: https://en.wikipedia.org/wiki/Time_complexity

วันพฤหัสบดีที่ 30 มกราคม พ.ศ. 2568

Normal vs Skewed vs Uniform distributions

Normal Distribution (Bell Curve): Characterized by a symmetric, bell-shaped curve, with the majority of data points clustered around the mean. This distribution is common in many natural phenomena, such as heights, weights, and IQ scores.

🕯 Skewed Distribution: Skewed distributions have a longer tail on one side compared to the other. There are two types:

💡 Right-skewed (Positive-skewed): The tail extends towards the higher values. It's also called positively skewed because the mean is greater than the median.

💡 Left-skewed (Negative-skewed): The tail extends towards the lower values. It's negatively skewed because the mean is less than the median.

🕯 Uniform Distribution: In a uniform distribution, all data points have equal frequencies, resulting in a flat histogram with evenly spaced bars. This distribution is commonly seen in random or equally likely events.

🕯 Bimodal Distribution: Bimodal distributions have two distinct peaks, indicating two separate groups or populations within the data.

🕯 Multimodal Distribution: Similar to bimodal but with more than two peaks, indicating multiple distinct groups or populations within the data.

วันพุธที่ 29 มกราคม พ.ศ. 2568

Reasoning model in LLM

What is a Reasoning Model in LLMs?

A reasoning model in LLMs refers to the mechanisms and processes that allow the model to:

Understand and process complex queries.
Draw inferences or conclusions from given information.
Solve problems by connecting concepts, facts, and patterns.
Generate coherent and contextually appropriate responses.

However, it's important to note that LLMs do not "reason" in the human sense. Instead, they simulate reasoning through pattern recognition and statistical associations learned during training.

Types of Reasoning Simulated by LLMs

LLMs can simulate various types of reasoning, including:

a. Deductive Reasoning

Drawing specific conclusions from general principles.
Example: "All birds can fly. A penguin is a bird. Can a penguin fly?" (LLMs can infer the correct answer based on learned knowledge.)

b. Inductive Reasoning

Making generalizations from specific examples.
Example: "Every crow I have seen is black. Therefore, all crows are likely black."

c. Abductive Reasoning

Forming the most likely explanation from incomplete information.
Example: "The grass is wet. It probably rained."

d. Analogical Reasoning

Drawing parallels between similar situations.
Example: "If the heart is like a pump, then the arteries are like pipes."

--Deepseek

O1 Pro (by 01.AI) --ChatGPT

Developer: 01.AI (a Chinese AI startup backed by Alibaba)
Release Date: 2024
Capabilities:
- Open-weight large language model (LLM)
- Trained with extensive datasets for high-performance NLP tasks
- Designed as a strong competitor to GPT-4-level models
- Supports multilingual processing and reasoning tasks

Key Differences

Feature	GPT-4o	O1 Pro (Yi-1 Pro)
Developer	OpenAI	01.AI (China)
Model Type	Proprietary	Open-weight LLM
Multimodal?	Yes (text, image, audio)	Primarily text-based
Performance	Advanced, real-time	Comparable to GPT-4
Accessibility	ChatGPT & API	Open-source weights available

วันอังคารที่ 28 มกราคม พ.ศ. 2568

MySQL CLI Tutorial

https://gist.github.com/hofmannsven/9164408

https://youtube.com/playlist?list=PLfdtiltiRHWEw4-kRrh1ZZy_3OcQxTn7P&si=pVet-eqarNaYcY6Q

วันอาทิตย์ที่ 26 มกราคม พ.ศ. 2568

Heart of machine learning

ML learns patterns from data with or without labels (i.e., supervised or unsupervised scheme).

Fuzzy logic doesn't learn from data. So it's not an ML.

วันพฤหัสบดีที่ 23 มกราคม พ.ศ. 2568

IEEE free data sets

https://ieee-dataport.org/

IEEE publishing ethics

https://journals.ieeeauthorcenter.ieee.org/become-an-ieee-journal-author/publishing-ethics/

Ethical requirements:

https://journals.ieeeauthorcenter.ieee.org/become-an-ieee-journal-author/publishing-ethics/ethical-requirements/#maincontent

วันจันทร์ที่ 20 มกราคม พ.ศ. 2568

NFT marketplace

https://opensea.io/

The world’s first and largest digital marketplace for crypto collectibles and non-fungible tokens (NFTs). Buy, sell, and discover exclusive digital items.

WebGPU

Newer and faster than WebGL