IntegrAuth Academy
Identity, security and practical AI — free, from zero. Byte-sized lessons, real diagrams, hands-on labs, zero fluff. Two programs, each with its own exam and certificate.
English is hard? Your browser can translate this page into Tamil, Telugu, Kannada, Malayalam or Hindi. Tech words stay in English so you learn the real names.
How?
- Chrome on phone: ⋮ menu → Translate
- Chrome on computer: right-click → Translate to …
- Edge: the translate icon in the address bar
What an AI model is
Priya's company just turned on an AI assistant. What is inside it? Your goal: see how a model writes, one word at a time.
🧪 Interactive lab — enable JavaScript to play with this one.
- A model is a huge file of numbers, plus a program that runs them.
- It writes by guessing the next piece of text, again and again.
- Training sets the numbers. Your chats do not change them.
- It knows nothing after its cutoff date, unless the app adds fresh facts.
One piece at a time
Your phone keyboard already does a small version of this. It suggests the next word as you type. An AI model is like that keyboard, after reading a huge library.
The model writes in small pieces. A piece can be a word, part of a word, or a comma. Each piece is a token. The model gives a chance to every possible next piece, like "login" 62%. One piece is picked and added. Then the loop runs again.
Training, then use
Inside are often billions of numbers, called weights. Training sets them. Each wrong guess moves every number a tiny bit. At the end, the numbers are frozen.
Using the finished model is called inference. Now the weights are only read. So your chats do not teach the model. It may seem to remember you. That is a separate memory feature that saves notes for new chats.
A frozen snapshot
Training text stops at a date. This date is the model's knowledge cutoff. The model knows nothing after it. So apps can add fresh facts to your question, like search results or a file. This is retrieval.
Now put a model's life in order.
🧪 Interactive lab — enable JavaScript to play with this one.
Go deeper: AI words
People often use these words as if they mean the same thing. They don't. Each one sits inside the one before it.
- Artificial intelligence (AI) is the whole field. Computers do tasks that seem to need a mind. It even includes programs where people wrote every rule by hand, like a route planner.
- Machine learning (ML) learns patterns from examples. Nobody writes the rules. A spam filter learned from emails that people marked as spam.
- Deep learning is machine learning with large networks of many layers. It made speech-to-text and face unlock work.
- Generative AI makes new content: text, images, audio, video or code. It does more than give a label like "spam."
- A large language model (LLM) is generative AI for language. Chat and coding assistants are built on LLMs.
So every LLM is AI, but most AI is not an LLM. The fraud check on your bank card is machine learning. It gives a risk score, not a paragraph.
Go deeper: the numbers inside
A model is a big chain of simple math. Numbers that stand for text go in. They pass through many layers. Then numbers come out. This is a neural network. Brain cells gave people the idea. But it is math, not a brain.
Each link between layers has its own number, a weight. The weights and a few other numbers are the model's parameters. GPT-2 small (2019) had 124 million. GPT-3 (2020) had 175 billion.
Think of a music mixing desk with billions of sliders. No single slider means "grammar" or "France." But set them all right, and the sound comes out right. A model file is a saved copy of every slider. Some makers let you download it. Others only let you use the model through their service. See Open and closed models.
Go deeper: two stages of training
Training has two stages. Pre-training comes first. The model reads a huge amount of text. It learns to predict what comes next. The result is a base model. It continues text, but it is not an assistant. Ask it a question, and it may just add more questions.
Post-training turns it into a helpful assistant. First it learns from example chats with good answers. Then people rank several answers to the same question. The model is pushed a little toward the ones they liked. This is called RLHF. Safety training teaches it what to refuse.
People often rate agreeable answers highly. So a model can learn to tell you what you want to hear. This is sycophancy.
Go deeper: Priya's leave days
Priya asks how many leave days she can carry over. The assistant says "five" and sounds sure. That was the rule until April. Then HR changed it to ten. Nothing was broken. The model wrote a likely answer from older text.
The fix was not a smarter model. IT connected the assistant to the current HR handbook. Now the new rule comes with each question.
Without a memory feature, a new chat starts fresh. It does not see your old chats.
A model has no built-in sense of what is true. It can sound sure and still be wrong. Asking "Are you sure?" only gets more likely text. Check a real source.
Transformer Explainer runs GPT-2 small, a real model, in your browser (the app you use to open websites). Watch the chances for the next token.