bharath.ai › Glossary
AI, explained through everyday India: a neural network as the Mumbai local, attention as a busy market, hallucination as an uncle’s confident directions.
Mumbai Local Train Network
Just as Mumbai’s complex train routes connect millions of people with decision-making at each junction, a neural network is a system of interconnected ‘neurons’ that process information layer by layer, learning patterns along the way.
Chai Wala’s Feedback Loop
A street chai walah adjusts his recipe based on customer feedback: more sugar for some, less for others. Similarly, backpropagation sends error signals backward through a neural network, adjusting the ‘recipe’ (weights) to improve predictions.
Learning from Experience
Like a cricket player improving through practice and experience rather than memorizing rulebooks, machine learning teaches computers to improve from data without explicit instructions for every scenario.
Cramming for Exams
A student who memorizes exam questions perfectly might fail when asked different questions on the same topic. Overfitting is when an AI model memorizes training data too well and fails on new, unseen data.
Finding Your Way Down a Hill Blindfolded
Imagine navigating down a dark hill by feeling the ground slope: you take small steps in the steepest direction. Gradient descent is how AI models ‘walk’ down the error landscape to find the best solution.
Breaking Hindi Into Words
Just as Hindi sentences break into meaningful words and phrases, tokenization breaks text into smaller pieces (tokens) that AI can understand and process individually.
Mapping People on a Train
If you map every person in a train by their position, wealth, and age in 3D space, words in AI work similarly: each word gets a unique position in a mathematical space where similar words are close together.
Focusing in a Crowded Market
In a bustling Indian market, your attention shifts to what’s relevant: a vendor’s call if you need produce, a sound if there’s danger. Attention in AI lets models focus on the most relevant information from vast input.
A Panchayat, Not a Relay Race
In a relay race only the runner holding the baton knows where the race stands, and older language models read that way, one word at a time. A transformer works more like a panchayat where everyone hears everyone at once: each word looks at every other word in parallel and weighs how much each matters to it. That parallel reading is what lets transformers learn from vast amounts of text quickly.
Adapting a Bollywood Script
A Bollywood script for Tamil cinema gets adjusted for local humor, references, and culture; the core story remains but details shift. Fine-tuning adapts pre-trained AI models for specific tasks.
A Pandit Who Read Every Book
An LLM is like a learned pandit who has read millions of texts and learned to predict what comes next in any conversation, having absorbed patterns from vast knowledge.
Asking a Google Voice Assistant
The way you ask ‘Hey Google, what’s the weather?’ vs ‘It’s raining outside, how should I dress?’ changes the answer. Prompt engineering is the art of asking AI precisely what you need.
The Indian Cricket Selection Committee
Like how India’s selection committee assigns different specialists for different conditions (pace bowlers for Australia, spinners for Sri Lanka), MoE routes inputs to specialized sub-networks based on the task. Only a few experts work on each input, so a very large model costs far less to run for every word it reads.
The Dharavi Social Network
In Dharavi, everyone knows everyone through connections: the leather guy knows the potter who knows the exporter. Knowledge graphs connect facts through relationships, so a system can answer a question such as which languages are spoken in which state by following the links.
Compressing a Bollywood Film for Mobile
Just as a 4K Bollywood film is compressed to play on a budget phone without losing the story, quantization stores a model’s numbers at lower precision, for example 4 bits instead of 16 or 32, so it fits and runs on affordable devices.
The Open-Book Exam
Instead of memorizing everything (like a closed-book exam), RAG lets AI look up relevant information from a knowledge base before answering, like a student who can check their notes during the test. It also lets the model point to the passage its answer came from.
The Overconfident Pandit
When an uncle at a wedding gives you confident, wrong directions to the venue, sure of every turn, that’s hallucination: AI generating plausible-sounding but factually incorrect information with complete confidence.
The Village Doctor vs City Hospital
Instead of sending every patient to AIIMS Delhi (cloud), an AI model running on a local device (edge) can diagnose basic conditions right in the village: faster, cheaper, and it works offline. It matters most where the network is weak or absent.