ඉතින් මේ Knowledge Graphs කියන concept එක අද කාලේ AI වල, විශේෂයෙන්ම "Retrieval-Augmented Generation" හෙවත් "RAG" systems වලදී නැතුවම බැරි දෙයක් වෙලා. RAG කියන්නේ ChatGPT වගේ Large Language Models (LLMs) වලට පිටතින්, up-to-date information දීලා, ඒවා "hallucinate" කරන එක (ඒ කියන්නේ බොරු දේවල් හදලා කියන එක) නවත්තන technique එකක් මේ ගැන අපි දීර්ඝ විදියට කතා කරා ඔයාලට මතක ඇතිනේ ඉතින් මේ සාමාන්ය RAG system එකක් කරන්නේ user ගේ ප්රශ්නයට අදාළව documents වලින් කොටස් (chunks) හොයලා LLM එකට දෙන එක. හැබැයි මේ chunks අතර සම්බන්ධයක් නැහැ. Knowledge Graph එකක් RAG වලට එකතු කරාම, මේ process එක ගොඩක් smart වෙනවා මොකද KG එකෙන් දෙන්නේ නිකම්ම text chunks නෙවෙයි structured කරපු, එකිනෙකට සම්බන්ධ වුණු facts. මේ නිසා LLM එකට ප්රශ්නයට අදාළ context එක සහා entities අතර තියෙන relationships හරියටම තේරුම් ගන්න පුළුවන් වෙනවා. මේක හරියට LLM එකට පොත් ගොඩක් දෙනවා වෙනුවට, හොඳට පිළිවෙලට හදපු විශ්වකෝෂයක් දෙනවා වගේ වැඩක්. මේ නිසා දෙන උත්තර ගොඩක් නිවැරදි වෙනවා, විශ්වාසවන්ත වෙනවා, සහ fact-based වෙනවා.

Knowledge Graph එකකින් කොච්චර complex දෙයක් වුණත් structure කරන්න පුළුවන් කියන එකට, අපි මේ image එකේ තියෙන example එක අරං බලමු. මේකෙ තියෙන්නේ පහුගිය දවස්වල social media එකේ වෙච්ච සිදුවීමක් වටා ගොඩනැගුණු Knowledge Graph එකක් සිද්ධිය දන්නවනේ 🤣. මෙතන "Watawala", "Dilmah", "Lipton" වගේ brands, "Advertisement", "LGBTQ+", "Straight Community" වගේ concepts ඔක්කොම "Nodes" විදිහට represent කරලා තියෙනවා. ඒ අතර තියෙන "supports", "created", "targeted", "triggered" වගේ සම්බන්ධතා "Edges" වලින් පෙන්නනවා. මේ Graph එක හැදිලා තියෙන්නේ social media posts එක එක sources වලින් "scrap" කරලා, ඒ කියන්නේ unstructured data අරගෙන.

ඉතින් මේ Graph එක දිහා බැලුවම අපිට එකපාරින්ම පේනවා "Watawala" brand එක "created" කරපු "Advertisement" එකක් "LGBTQ+" community එක "promotes" කරන අතරේ, "Straight Community" එක "triggered" කරලා තියෙනවා කියලා. ඒ වගේම, මේකට ප්රතිචාර විදිහට "Watawala" එකෙන් "Apology" එකක් නිකුත් කරලා තියෙනවා. මේ වගේ සංකීර්ණ, මතභේදාත්මක මාතෘකාවක් ගැන පැහැදිලි, structured view එකක් AI එකකට දෙන්න Knowledge Graph එකකට පුළුවන් නිකම්ම text එකක් කියවනවාට වඩා, මේ වගේ graph එකක් හරහා AI එකකට සිද්ධියේ සම්පූර්ණ context එක, ඒකට සම්බන්ධ පාර්ශ්ව, සහා ඔවුන් අතර ඇති සම්බන්ධතා හරියටම තේරුම් ගන්න පුළුවන් වෙනවා. මේක තමයි Knowledge Graph එකක ඇත්තම power එක.

#ZuuCrew #Ai

✍️ @Zuu Crew (LinkedIn)

Cyber war 😂

📊 Model Evaluation — Image Understanding Leaderboard (2025)

AI models are ranking based on how they interpret and understand images from distribution applications across two key tasks like generating detailed captions and assessing image quality.

OpenAI is still leading the AI landscape. 🚀

#ImageAI #VisionModels #AIEvaluation #ModelBenchmark #ImageUnderstanding #ComputerVision #AIResearch #DeepLearning #TechNews #OpenAI #FutureOfAI #AILeaderboard #InfinityAI

✍️ @TheInfinityAI

Wooh 🤓

🧠 Understanding State Management in React.js

when we are building modern web apps with React.js, the state management is one oof the most important concept. it is the foundation that allows apps to update and react to user interactions and data changes.

🌱 What Is "State" in React?

No need to go deep. in simple words, State is data that is changing over time with time and it affects the content of the screen that user sees.

⚛️ The Two Types of Data in React

Props - data passed in to a component from its parent.
State - data managed inside a components

Props are like inputs and State is like the memory of a component.

🪄 How to Manage State

useState() hook in React used to create and mange local state inside a functional component.

import React, { useState } from "react";

function Counter() {
  const [count, setCount] = useState(0); // Initial state = 0

  return (
    <div>
      <h2>Counter: {count}</h2>
      <button onClick={() => setCount(count + 1)}>Increase</button>
    </div>
  );
}

export default Counter;

When the user clicks Increase, React updates the state and automatically re-renders the component and showing the new count to the user.

🔁 Why State Management Matters

When your app is growing, also growing the complexity of managing states

● Data may need to shared between multiple components.
● Some data might come from API or a database.
● When need to keep the UI synchronized with real-time updates.

Thats where state management becomes important, keeping your app’s data organized, predictable and easy to control.

🧩 Types of State in React

React applications deal with four main types of state

1. Local State - State that belongs to one component only.
2. Global State - Data that multiple components need to access.
3. Server State - Data fetched from a backend server or API.
4. URL State - Information stored in the browser’s URL

🛠 Common Tools for Managing State

● React Context API
● Redux Toolkit
● React Query (TanStack Query)

#ReactJS #StateManagement #ReactHooks #WebDevelopment #FrontendDev #JavaScript #useState #ReactTutorial #CodeNewbie #WebDevCommunity #LearnReact #ProgrammingTips #TechEducation #InfinityAI

✍️ @TheInfinityAI

🔥 Big news for multilingual NLP! Check out this new paper on MMBERT

🌐🤖 Summary: The PDF describes how a modern encoder model (MMBERT) was trained using massive multilingual data. It shows how it achieves better results in language understanding and translation tasks across many languages.

#MMBERT #MultilingualNLP #NLPResearch #AIModels #DeepLearning #BERT #LanguageAI #MachineLearning #ArtificialIntelligence #NLPCommunity #MultilingualAI #TechNews #InfinityAI

🚀 Free Python Full Course (සිංහල Medium)
🔹 Valid Until: November 06
🎓 Beginner Friendly | No Prior Experience Required

මේ Course එකෙන් ZERO to Advanced Python Programming පියවරෙන් පියවර ඉගෙනගන්න පුලුවන්. Variables, Data Types, Operators, Loops, Functions, OOP, Databases, GUI Development වැනි මුලික + ප්‍රායෝගික දේවල් හොදින් explain කරලා තියෙනවා.

💡 Ideal For:
• Programming beginners
• AI / Data Science / Automation learners
• Daily task automation with Python

📚 Course Duration: 15+ Hours
📝 Language: Sinhala
🧠 Simple Examples + Clear Explanations

🔗 Course Link:
https://www.udemy.com/course/python-for-everyone-sinhala-medium

🎁 FREE Access Coupon Code: CODEPROLKPYTHON
⏳ Use Before: November 06
——————————————————————
🔥 Grab It Fast & Save to Your Udemy Library
👨‍💻 Python is Must-Learn Today — Start Now!

🎉 Special Thanks to CodePRO LK for providing this amazing free learning opportunity!
👉 @codeprolk (YouTube)

✍️ @TheInfinityAI

Understanding the Transformer Model in LLMs

Introduction

When you are using ChatGPT, Google Gemini, Claude or any other modern AI Language Model, you are interacting with a transformer. The transformer architecture introduced in 2017 by Google researchers revolutionized the ability of computers to understand and generate human language. This article will walk you through what transformers are and how they work step-by-step. Before transformers, language models looked at one word at a time. It was like reading a book with a flashlight that only shows one word. This made it hard for the model to understand long sentences. As an Example, In the sentence "The bank was steep," the word "bank" could mean a financial bank or a riverbank. To choose the right meaning, the model needs to see the word "steep". Transformers fix this by looking at all words in a sentence at once. This helps the model see how each word connects to the others

Step 01 - Breaking Text Into Tokens

Before the transformer reads the text, it splits it into smaller units called tokens. Each token becomes a number (a vector) that the model understands. This is called tokenization and embedding.
- Input - "I love artificial intelligence"
- Tokens - ["I", "love", "art", "ificial", "intell", "igence"]

Step 02 - Adding Positional Information

Transformers look at all words at once, so they need to know the order. For example, "dog bites man" is very different from "man bites dog." Positional encoding adds order to the tokens. It's like giving page numbers to mixed-up book pages.

Step 03 - Attention Mechanism

This is the key part of transformers. Attention helps the model focus on important words when reading a sentence. Let's try to understand how Attention Works. Let's assume we need to translate "Le chat noir" (the black cat) into English. For this, the model needs to focus on "Le" and "noir" when working on "chat." The model uses three parts for this

01. Query - What word are we looking at now?
02. Key - What other words are there?
03. Value - What meaning do they give?

The model gives attention scores to all words to decide which ones are more important. Transformers use many attention heads at the same time (Multi-Head Attention). Each head looks at different things like grammar, meaning or logic. This helps the model understand better.

Step 04 - Feed-Forward Neural Networks

After attention, the model sends each word's information through a feed-forward neural network. This helps the model learn more and improve its guesses. It’s like asking, “Now that I understand the sentence better, now what else can I figure out about this word?”

Step 05 - Stacking Layers

Transformers use many layers, not just one. Each layer improves the output of the one before it.

Layer 1 - Finds simple word links
Layer 5 - Understands grammar better
Layer 10 - Handles deep meaning
Layer 20+ - Deals with complex ideas and reasoning

It’s like editing a paper many times to make it a better and clearer output.

Step 07 - Training the Model

Transformers learn by reading lots of text. During training:

- The model reads billions of sentences
- It guesses the next word
- If wrong, it makes small changes to improve
- It repeats this many times

This helps the model learn grammar, facts, patterns, and some logic.

Step 8 - How It Writes Text

When you ask a question:

- Your text becomes tokens
- Tokens go through all the layers
- Attention helps the model find meaning
- It guesses the next word
- Adds it to the answer

This will repeat until the answer is complete

Conclusion

Transformers are more powerful because they can read all words at once, understand long texts well, work better with more data and can be used for many tasks. Transformers changed how AI understands language by focusing on what matters and using many layers of learning. It became the core of most modern language models. They are not magic, but they are clever tools that learn from huge amounts of text and help computers understand and create human language better than ever before.

✍️ @TheInfinityAI

🎉 ChatGPT Go Plan – Now Free for 12 Months in India! 🇮🇳

OpenAI is offering the Go plan free for one year to users in India. This plan unlocks faster image generation, expanded file uploads, better memory, and longer context — all powered by GPT-5!

🌍 Live outside India? No problem!

You can still access this amazing offer using an Indian VPN. Just connect, refresh your browser, and upgrade to Go at ₹0/month for the first 12 months.

✨ Don't miss this limited-time opportunity to explore advanced AI features without paying a single rupee!

👉 Yeh, Try it now and supercharge your ChatGPT experience.

#ChatGPTGo #ChatGPTUpdate #OpenAI #AIIndia #GPT5 #TechNews #AIOffers #VPN #AICommunity #ArtificialIntelligence #AIFeatures #InfinityAI

@TheInfinityAI

OpenAI has officially rolled out GPT 5.1 worldwide. Paid users receive access first, with free users expected to join soon.

This release introduces two upgraded models. GPT 5.1 Instant delivers a warmer, more natural tone and follows instructions more smoothly. GPT 5.1 Thinking focuses on deeper reasoning, handling simple questions faster while dedicating more time to complex tasks.

What’s new:

▪️Adaptive reasoning that automatically adjusts thinking time
▪️Tone and personality presets such as Friendly, Professional, and Quirky
▪️Improved performance in math, coding, and logic evaluations
▪️Smoother conversation flow and better contextual memory

#OpenAI #GPT5 #ChatGPTUpdate #AIFeatures #TechNews #ArtificialIntelligence@TheInfinityAI

The biggest change in this new model is that it’s built to work like a "Digital Employee," not just a regular Chatbot.

Anthropic has introduced three main features this time: Effort Parameter, Context Compaction, and Tool Search. Usually, when we ask an AI a question, it just gives a standard answer. But with the Effort Parameter in Opus 4.5, we can decide how long the model should think. If you set it to "High Effort," it thinks much deeper than the old Opus 4.1 to solve coding problems. The amazing thing is, even at "Medium Effort," it performs just as well as the previous Sonnet 4.5 but uses 76% fewer output tokens. That’s a huge win for saving money.

The next most important thing is Context Compaction. We know that with any AI, the "memory" fills up as you keep working, and it forgets what you said at the start. But Opus 4.5 compresses old logs and unnecessary details, keeping only the important stuff in its active memory. Because of this, it can handle coding sessions that last for 30 hours straight. Also, thanks to Tool Search, it can now handle thousands of enterprise tools. It has the ability to search for and find the right tool when needed, just like a human looking up a manual to get work done.

Talking about Benchmarks, Opus 4.5 scored 80.9% on the SWE-bench Verified, which is the industry standard. This is a significant lead over Google Gemini 3 (76.2%) and GPT-5.1 (~76.3%). This isn't just a random number. It shows that Opus 4.5 can now handle bug fixing and feature implementation on its own, just like a Junior Developer. For example, it managed to fix a complex bug in a PostgreSQL extension—something previous models couldn't do—by cloning the repository and checking the internals all by itself.

On the Pricing side, Anthropic has made a big move. The old Opus 4.1 was expensive (around $15 input / $75 output). But the new Opus 4.5 is about 66% cheaper ($5 input / $25 output). Even though it's still a bit pricier than GPT-5.1 ($1.25/$10), Opus 4.5 saves you money in the long run because of its "First-pass reliability." Since it writes the correct code the first time, you don't waste money fixing errors.

In short, Claude Opus 4.5 isn't just another AI update. While Google is focusing on video and fancy features, Anthropic has chosen to focus on Pure Intelligence and Reliability. We can confidently say this is the best tool right now for complex work like Software Engineering and Data Analysis. It’s clear now that by 2026, the world of enterprise software is going to change completely because of these "Agentic" AI models.

@TheInfinityAI

ඔයා දන්නවද Uncensored LLMs ගැන, අපි හැමෝම වගේ ChatGPT, Claude එහෙමත් නැත්නම් Gemini වගේ AI tools පාවිච්චි කරන්න පුරුදු වෙලානේ ඉන්නේ. හැබැයි ඔයාට කවදා හරි හිතිලා තියෙනවද මේවා හිර වෙලා ඉන්න රාමුවෙන් එළියට ගියොත් මොන වගේ වෙයිද කියලා? අන්න එතනදි තමයි මේ Uncensored LLMs කියන මාතෘකාව කරළියට එන්නේ.

සරලවම කිව්වොත්, මේ මාදිලි හරියට 'බෑ' කියන්න දන්නේ නැති AI එකතුවක් වගේ. දැන් බලන්නකෝ, අපි සාමාන්‍ය ChatGPT එකෙන් ටිකක් මතභේදාත්මක හෝ අනතුරුදායක දෙයක් ගැන ඇහුවොත්, ඒක බොහොම ආචාරශීලීව ඒ ඉල්ලීම ප්‍රතික්ෂේප කරනවානේ. ඒකට හේතුව තමයි ඒවාට දාලා තියෙන Safety Guardrails නැත්නම් ආරක්ෂක වැටවල්. හැබැයි මේ Uncensored models වලට එහෙම කිසිම තහංචියක් නෑ. ඕනෑම දෙයක්, ඕනෑම විදියකට කතා කරන්න මේවා ලෑස්තියි. මේක එක පැත්තකින් භයානකයි වගේ පෙනුනට, පර්යේෂණ කරන අයට සහ නිර්මාණාත්මක දේවල් ලියන අයට නම් මෙය මාරම නිදහසක්.

මේවා නිර්මාණය කරන තාක්ෂණයත් හරිම පුදුම හිතෙන සුළුයි. මේ දවස්වල, ඒ කියන්නේ 2025 වෙද්දි ගොඩක් ජනප්‍රිය වෙලා තියෙන්නේ Abliteration කියන ක්‍රමවේදය. ඒ කියන්නේ අර Llama 3 වගේ ප්‍රබල මොඩල් එකක් අරගෙන, Model එකේ තියෙන refusal mechanism එක (ඒ කියන්නේ එපා කියන්න බල කරන කොටස) mathematically අයින් කරන එක. හරියට නිකන් අපේ මොලේ ඔපරේශන් එකක් කරලා බය කියන දේ අමතක කරවනවා වගේ වැඩක්.

දැන් ඔයාට ප්‍රශ්නයක් එන්න පුළුවන්, මොනවද මේ වෙද්දි තියෙන හොඳම Models කියලා.

Code ලියන්නයි, තර්කානුකූලව හිතන්නයි නම් මේ වෙලාවේ හැමෝම කතා වෙන්නේ Dolphin 3.0 එක ගැන. ඒක Llama 3.1 එක පදනම් කරගෙන හදපු හින්දා හරිම වේගවත් වගේම තියුණුයි. අනික නිර්මාණාත්මක දේවල් ලියන්න කැමති අයට Llama-3.2 Dark Champion එක ගොඩක් හොඳයි.

Images එහෙම හදන්න කැමති නම්, Midjourney එකටත් වඩා ලොකු නිදහසක් Flux.1 එකෙන් ලැබෙනවා. මොකද මේකේ Midjourney වගේ 'Banned Words' ලිස්ට් එකක් නෑ. හිතේ තියෙන ඕනම අදහසක් රූපයක් කරන්න මේකට පුළුවන්. ඒ වගේම දැන් වීඩියෝ හදන අයටත් Wan 2.1 සහ Open-Sora වගේ tools ලොකු වාසියක් වෙලා තියෙනවා, මොකද මේවා ඕනම කෙනෙක්ට තමන්ගේ PC එකේම රන් කරන්න පුළුවන් නිසා.

විශේෂම දේ තමයි, අපේ සිංහල භාෂාව හසුරුවන්න පුළුවන් JOSIEFIED-Qwen3 වගේ මාදිලිත් මේ අතර තියෙන එක. ආසියානු භාෂා එක්ක වැඩ කරන්න මේක හරිම දක්ෂයි.

කොහොම වුණත්, මේ තාක්ෂණය හරිම risky. මේකෙදි අපිට ලැබෙන නිදහස එක්කම ලොකු වගකීමකුත් එනවා. මොකද මේක පාවිච්චි කරන කෙනාගේ අරමුණ මත තමයි ඒක තීරණය වෙන්නේ.

@TheInfinityAI