Generative AI refers to AI systems that can create new and original content — text, images, audio, video, and code — that resembles human-created content. It does this by learning patterns and structures from vast datasets using machine learning and deep learning models.
💡 Think of it This Way
Traditional AI is like an examiner — it reads your answer and gives you a score. Generative AI is like a student — it reads lakhs of answers and then writes its own original answer. That is the fundamental shift: from analysis to creation.
Hallucinations — generates confident but factually wrong outputs (risky in healthcare/law)
Scalability — serves large populations simultaneously
Erosion of human creativity and critical thinking through over-reliance
Decision support — summarises large datasets, generates actionable insights
Digital divide — benefits concentrated among those with data and computing access
Generative AI vs Traditional AI
Feature
Traditional AI
Generative AI
Primary Function
Analyses existing data
Generates new content
Nature of Output
Predictions, classifications, decisions
Text, images, audio, video, code
Output Behaviour
Deterministic or rule-based
Probabilistic and creative
Learning Approach
Rule-based or predictive models
Deep learning-based generative models
Data Requirement
Moderate to large datasets
Very large datasets
Explainability
Relatively better
Low (black-box)
Risk Profile
Bias, prediction errors
Deepfakes, misinformation, hallucinations
Examples
Spam filtering, fraud detection, face recognition
Chatbots, image generators, code assistants
Large Language Models (LLMs)
Large Language Models are advanced Generative AI models trained on massive text datasets to understand, generate, and manipulate human language with high accuracy. They are built primarily using deep learning and transformer architectures.
🗣️What Makes an LLM ‘Large’?
Think of parameters as the ‘knowledge capacity’ of a model. GPT-3 had 175 billion parameters. GPT-4 is estimated to have over 1 trillion. More parameters trained on more data = better language understanding and generation. The ‘large’ in LLM refers to the scale — of data, computation, and parameters.
Key Characteristics of LLMs
Context Awareness: Can understand and maintain context over long text passages — enabling coherent conversations, summaries, and multi-paragraph reasoning
Scalability: Performance improves with larger data and more parameters
Generative Capability: Produces fluent, human-like text for writing, translation, summarisation, and Q&A
Multilingual Ability: Processes and generates multiple languages — supporting cross-lingual understanding and inclusive digital services
Types of LLMs
Classification
Type
Examples & Notes
By Architecture
Encoder-only Models
Focus on understanding and analysing text (not generating). Used for classification, sentiment analysis. E.g., BERT.
By Architecture
Decoder-only Models
Designed for text generation — predicts next word based on context. Most popular LLMs today. E.g., GPT, LLaMA.
By Architecture
Encoder-Decoder Models
Combines understanding + generation. For translation, summarisation, Q&A. E.g., T5.
By Openness
Closed/Proprietary LLMs
Controlled by private companies; limited access. Offered via APIs. E.g., GPT, Claude, Gemini.
By Openness
Open/Open-Weight LLMs
Model parameters publicly available; encourages innovation, transparency. E.g., LLaMA, Falcon.
By Capability
Text-only LLMs
Process and generate text only. E.g., early GPT models.
By Capability
Multimodal LLMs
Process text, images, audio, video, code. E.g., Gemini.
By Usage
General-Purpose LLMs
Perform multiple tasks across domains. E.g., GPT, Claude.
By Usage
Domain-Specific LLMs
Fine-tuned for law, healthcare, finance, scientific research.
LLMs vs Traditional NLP vs Chatbots
Feature
LLMs
Traditional NLP
Chatbots
Nature
Generative AI models
AI sub-field, task-based
Applications/interfaces for conversation
Architecture
Transformer-based deep learning
Rule-based, statistical, ML/DL
Rule-based or LLM-powered
Task Handling
Multi-task (one model, many tasks)
Single-task or pipeline-based
Task-specific or domain-specific
Context Understanding
Strong (long-range)
Limited context handling
Limited (often short, predefined flows)
Output
Probabilistic, generative
Mostly analytical/classificatory
Scripted or generative if LLM-powered
Explainability
Low (black-box)
Relatively better
Varies
Compute Needed
Very high
Low to moderate
Low to moderate
Examples
GPT, Claude, Gemini, LLaMA
POS tagging, NER, Sentiment
Customer support bots, FAQ bots
Types of Chatbots
Type
Description
Rule-Based Chatbots
Operate on predefined rules and decision trees. Respond only to specific commands. No learning capability. E.g., FAQ bots.
ML-Based Chatbots
Use ML to improve responses from past interactions. Limited contextual understanding. E.g., basic customer service bots.
NLP-Based Chatbots
Use NLP to understand user intent and language variations. More flexible than rule-based bots. Handle free-text inputs.
LLM-Powered Chatbots
Use LLMs for context-aware, generative, human-like conversations. Can perform multiple complex tasks. E.g., ChatGPT, Gemini.
Hybrid Chatbots
Combine rule-based logic with AI/LLM intelligence. Rules handle routine queries; AI handles complex ones.
Voice-Based Chatbots
Interact through speech using speech recognition, NLU, and AI. E.g., IVR systems, voice assistants like Alexa.
Important Global Large Language Models
LLM
Developer
Key Features
GPT Series
OpenAI
Benchmark frontier LLM; strong reasoning; multilingual; widely used globally
Gemini
Google DeepMind
Natively multimodal; successor to PaLM; integrates text, image, audio, video
Claude
Anthropic
Safety-first LLM using Constitutional AI — aligned with human values
LLaMA
Meta
Open-weight model; backbone of open-source and global LLM ecosystems
PaLM
Google
Legacy foundation model; precursor to Gemini
Mistral/Mixtral
Mistral AI
Europe’s push for AI sovereignty; efficient and open-weight
The Internet is the single most impactful invention of the 20th century — more impactful than the automobile or the television, because it contains and accelerates all other technologies. The Internet is a global network of interconnected computers and communication systems that allows users to share information, access services, and communicate across long distances. It…
Space Debris: The Scattered Danger in Orbit Let’s begin with a small everyday picture. Imagine you are driving on a highway at 100 km/h, and suddenly a tiny pebble flies up and hits your windscreen. Even at that modest speed, a small stone can crack the glass. Now take that same scene and move it…
Militarisation of Space There is an important conceptual distinction you must remember: Militarisation vs. Weaponisation. Militarisation is using space to SUPPORT military operations — like using GPS for troop coordination. Weaponisation is placing or using WEAPONS in space itself. The Outer Space Treaty (1967) does not ban militarisation. It only bans WMDs in orbit. This…
Space Traffic Management (STM) Think of outer space as a busy international airspace — except there is no Air Traffic Control, no unified radar network, no binding rules for who gives way to whom, and no international police. Over 9,000 active satellites and millions of debris fragments share the same sky. Without STM, catastrophic collisions…
Major Lunar Missions — The Story of the Moon The Moon has always been humanity’s first space destination — simply because it is the nearest. But getting there was never a straight line; it was a clear evolution: first we went just to take photographs, then we attempted soft landings, then we sent humans, and…
India is not a latecomer to science. Before Europe had universities, India had Nalanda and Takshashila. Before the West developed algebra, India gave the world Zero. Before modern surgery, India had Sushruta performing rhinoplasty. The story of Science and Technology in India is, in many ways, the story of a civilisation’s journey — from ancient…
