AI Terms Explained with Simple Examples Using a Dosa Analogy

In this blog post, we will see AI terms explained with simple examples, covering the buzzwords that keep flying around these days (artificial intelligence, machine learning, tokens, RAG, GPU, and a lot more). An engineer explains common AI terminology by mapping each concept to the workings of a dosa kitchen, where the chef, tawa, recipe notebook, and order counter serve as analogies for models, GPUs, parameters, and APIs. Terms covered include LLMs, RAG, tokens, inference, fine-tuning, agents, and hallucination. The analogies are designed to make abstract concepts accessible to technical professionals encountering AI terminology in workplace settings. Each term is paired with a practical use case, such as RAG powering an HR chatbot or a coding agent fixing a failing test suite. Lately every standup, every LinkedIn post, and every client call has at least one AI term flying past. I found myself nodding along in a few meetings without fully knowing what “tokens per second” or “mixture of experts” actually meant. So I sat down, broke each term apart, and mapped it to something we all understand well: making dosa. Dosa is one of the prominent south Indian breakfast 🙂 Yes, dosa. Stick with me, this actually works surprisingly well. Table of Contents Toggle The Dosa Cloud Kitchen setup Imagine a chef who makes dosa. Not just one dosa, thousands of them, for thousands of customers, all day. That chef, the kitchen, the tawa, the recipe notebook, the ingredient store, all of it together is our analogy for how AI systems work. Every term below maps to one small piece of this kitchen. By the end, the whole picture should click. The big picture: AI, ML, LLM Artificial Intelligence (AI) AI is the umbrella term for any system that can make decisions or perform tasks that normally need human intelligence. In our kitchen, this is simply “a smart cook who can figure things out” instead of a robot arm that only flips dosa the exact same way every time with no judgment. Machine Learning (ML) ML is how we get that smart cook without writing a thousand rigid rules like “if batter is this thick, then pour this much.” Instead, we show the cook thousands of dosa examples (good ones and bad ones) and let them learn the pattern themselves. That is machine learning: learning from examples instead of hardcoded instructions. Large Language Model (LLM) An LLM is a specific kind of ML system trained mainly on text, so it is very good with language. Think of it as a chef who did not just learn dosa making, they read every cookbook, food blog, and recipe card ever written, and now they can discuss, explain, and generate text about food (or almost anything) fluently. ChatGPT, Claude, and Gemini are all LLMs. Use case: a support chatbot that reads a customer’s question and replies in natural language is powered by an LLM, not a simple rule-based script. The brain of the kitchen Model The model is the trained chef itself, the end result after learning. It is not the ingredients, not the kitchen, just the chef’s skill and knowledge, packaged up so it can be reused anywhere. Parameters Parameters are the countless tiny dials the chef adjusted while learning: how much batter to pour, how long to wait before flipping, how much oil to use. A model with more parameters (like 70 billion vs 7 billion) has learned more of these fine dials, which usually means better judgment, but also a bigger, heavier chef to carry around. Training Training is the learning phase itself, the weeks the chef spent watching dosa after dosa being made, adjusting their technique each time they got it wrong. This happens once, offline, before the chef ever serves a real customer. Training Data Training data is everything the chef studied to learn: thousands of dosa recipes, videos, photos, customer reviews. The quality of the chef directly depends on the quality and variety of this data. Feed the chef only burnt dosa examples, and guess what they will confidently serve you. Fine-tuning Fine-tuning is a short, focused refresher course after the chef already knows the basics. Say the chef already knows all of Indian cooking, but a restaurant wants them to specialize only in Chettinad-style dosa. Instead of retraining from scratch, they get a short, targeted course just on that style. That is fine-tuning, cheaper and faster than full training. Talking to the chef Prompt A prompt is simply your order to the chef: “make me a masala dosa, extra crispy, less oil.” How clearly you phrase the order affects how good the result is. Vague orders get vague dosa. Tokens Tokens are the small chunks of text the chef reads and writes, roughly pieces of words rather than whole sentences. When you type “crispy masala dosa,” the chef does not see it as one block, it sees small bite-sized pieces, a bit like how a chef reads a recipe card word by word and sometimes syllable by syllable, not the whole page in one glance. Context Window The context window is how much of your conversation the chef can hold in their head at once, like the counter space in front of them. A small counter means they forget the beginning of your order once you add too many extra requests. A bigger context window means they can remember your entire order history, including that you mentioned an allergy five messages ago. Tokens per Second This measures how fast the chef can produce output, literally how many word-pieces they can write out per second. A chef serving tokens fast feels snappy and responsive. A slow one makes you stare at a loading dot wondering if your order got lost. This is a direct measure of speed and user experience for any AI product. Serving the customer Inference Inference is the actual moment of cooking, when the trained chef takes your real order and produces a real dosa on the spot. Training happened once in the past; inference happens every single time a customer orders. Inference Provider Not everyone wants to own a kitchen. An inference provider is a company that already has the trained chef and the tawa ready, and rents out cooking time to you. You place an order, they cook, you pay per dosa. Groq, Together AI, and AWS Bedrock are examples of inference providers for AI models. API The API is the order counter, the menu, and the ticket system combined. You do not walk into the kitchen and mess with the tawa yourself, you place your order at the counter in a format the kitchen understands, and a dosa comes back out. That interface is the API, it lets your application talk to a model without knowing anything about what happens inside the kitchen. The equipment GPU A regular stove (CPU) can make one dosa reasonably well. A GPU is like a massive industrial tawa with hundreds of small hotplates side by side, built to cook many dosas in parallel instead of one at a time. AI models need this kind of parallel firepower because they are doing millions of tiny calculations at once, not one long calculation. Data Center A data center is the giant central kitchen facility housing hundreds of these industrial tawas (GPUs), all networked together, kept cool, and powered around the clock. Companies like Nvidia, Microsoft, and Google run these at massive scale to serve everyone’s dosa orders simultaneously. Quantization Quantization is compressing the chef’s detailed recipe notebook into a smaller pocket-sized version. You lose a bit of the finer detail (maybe an exact gram measurement becomes “a small handful”), but the notebook now fits in a home kitchen instead of needing an industrial one. This is how large models get squeezed down to run on smaller hardware, sometimes even a laptop, with a small tradeoff in precision. Sharing recipes Open Weight Open weight means the chef publishes their full recipe notebook (the learned parameters) for anyone to download and run in their own kitchen. It does not always mean they share the raw training data or the exact steps of how they trained, just the finished, usable notebook. Meta’s Llama models are a well-known open weight example. Mixture of Experts (MoE) Instead of one chef trying to master every dish, imagine a kitchen with specialist chefs: one only does plain dosa, another only does masala, another only does uttapam. A head chef looks at each order and routes it to the right specialist instead of making everyone cook every order. This means only a portion of the whole team works on any given dish, which saves effort while still keeping expert-level quality. Models like Mixtral and DeepSeek use this mixture of experts approach. Hugging Face Hugging Face is like a giant online market where chefs from all over the world publish their recipe notebooks (models) and datasets for others to browse, download, and use. Head to huggingface.co and you will find thousands of ready-made “chefs” for almost any task, from dosa-writing to code-writing. Looking things up RAG (Retrieval Augmented Generation) A chef cannot memorize every regional variation of dosa ever invented. RAG is the chef quickly flipping through a filing cabinet of recipe cards to look up a specific one right before cooking, instead of relying purely on memory. This is how AI systems stay accurate about very specific or recent information, by fetching relevant documents first, then answering based on what they found. Use case: a company chatbot that answers questions about internal HR policy uses RAG to pull the exact policy document before replying, instead of guessing from general training. Vector and Embeddings To search that filing cabinet quickly, every recipe card is converted into a set of numbers that represents its “flavor profile,” so similar dishes end up near each other numerically. That numeric representation is called an embedding, and the space it lives in is called a vector space. This is how a system finds “recipes like this one” in milliseconds instead of reading every card one by one. Where you cook Hosting Hosting is simply where the kitchen physically operates: rented cloud kitchen space (a data center run by AWS, Azure, GCP) or your own home kitchen. Where you host affects cost, control, and speed. Local AI Local AI is cooking the dosa yourself at home instead of ordering from a restaurant. You need your own tawa (GPU or capable hardware), but you get full privacy since no order ever leaves your kitchen, and no per-dosa fee to a provider. Tools like Ollama let developers run models like Llama locally on their own machine, popular for privacy-sensitive or offline use cases. Money talk Model Cost Model cost is the price per dosa, usually measured per million tokens in and out. A bigger, smarter chef (larger model) generally charges more per dosa than a smaller, faster one. Picking a model is often a tradeoff between quality of the dosa and the bill at the end of the month, exactly like choosing between a five-star restaurant and a quick roadside stall. People behind it Forward Deployed Engineer This is the engineer who does not just build the kitchen equipment back at headquarters, they travel out to the actual customer’s restaurant, sit with their staff, and customize the whole setup to that specific menu and workflow. In AI companies, a Forward Deployed Engineer works directly with a client to integrate and tailor AI systems to that client’s exact data and processes, rather than shipping a one-size-fits-all product. The chef builds a team Agent A regular chef waits for your order, cooks it, and stops. An agent is a chef who takes initiative: checks what ingredients are running low, places a reorder, decides today’s special based on what is fresh, cooks it, and updates the menu board, all without you asking for each individual step. An agent decides the next step on its own instead of waiting to be told each move. Agentic AI Agentic AI is the general label for systems built around this kind of behavior, ones that plan, take multiple s [truncated for AI cost control]