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
Cheat sheet & pop quiz
You finished the track. Here it is on one page, plus 5 quick questions. Your goal: answer each one before you reveal it.
If you remember nothing else
| Lesson | The idea to keep |
|---|---|
| What an AI model is | A model is a next-token guesser. Training sets its weights, and then they are only read. It does not learn from your chats. It knows nothing after its knowledge cutoff. |
| Tokens and context windows | Everything is counted in tokens. Input and output share one context window. Output costs more. Details buried in a long text are easy to miss. |
| Inside one answer | Reading the prompt is one pass. Writing is one pass per token, so output is the slow part. Saved notes (the KV cache) make prompt caching work. A mixture of experts is big in memory but does little math per token. |
| Why answers change: temperature | Temperature sets how much answers vary. Use it low for facts and code, higher for new ideas. It changes variety, not truth. |
| When AI makes things up | Sounding sure is not being right. A citation is a claim to check. Give it sources. Ask neutral questions, so it does not just agree with you. |
| Prompts that work | A prompt is a brief: goal, context, format and a few examples. Let it say "not sure". The system prompt is not a security wall. |
| Choosing a model | Match the model to the job: size, effort, speed and cost. Send easy work to small models. |
| Models that pick from a list | A decision model picks from a fixed list in one fast step. It can only pick options you gave it. Add "other", and let a person decide when its confidence is low. |
| Open and closed models | Open weights are not open source. Read the license, not the label. A smaller, quantized model can run on your own computer. |
| Search by meaning: embeddings and RAG | Embeddings turn text into arrows. Arrows that point the same way have a similar meaning. RAG finds the right pages and adds them to the prompt. |
| Prompt, RAG or fine-tune? | Cheapest fix first: prompt, then RAG for facts that change, then fine-tuning for habits. Never fine-tune for facts. Exact jobs are for plain code. |
| Evals: does your AI work? | Evals, not vibes. Use a fixed set of real cases and a way to score them. Check case by case after every change. |
Pop quiz: 5 questions
Q1 · Priya asks the company assistant for this year's new expense limit. It answers with confidence, but gives last year's number. Nobody changed the model. What happened? What is the fix?
The model answered from frozen weights. Its training text is older than the change. It does not look things up, and it does not learn from chats. So it gave a likely answer. It did not say "I don't know." The fix: give it the current policy with the question. Attach it, or connect retrieval. Then ask it to quote the part it used. See What an AI model is.
Q2 · A help bot sends a 40-page manual and the whole chat with every message. It is slow and costly. It also keeps missing a rule on page 22. What is going on? What would you change?
It is all tokens. Everything sent again counts as input tokens on every turn. A crowded window is slower, and a rule buried in the middle is easy to miss. Send only the parts that matter, with retrieval. Start fresh chats, or carry a short summary instead of the whole chat. See Tokens and context windows.
Q3 · The same support ticket gets a different label each time. Meanwhile, the marketing team wants "more creative" product names from the same model. What do you change for each? What will it not fix?
For labels, lower the temperature. Also give a fixed list of labels, with an example of each. For names, raise the temperature, if the model allows it. Neither makes the model more correct. Temperature changes variety, not truth. Labels can be the same every time and still be wrong. So keep checking a sample. See Why answers change: temperature.
Q4 · The legal team wants AI help with contracts. The contracts must not leave the building. Someone says, "Let's fine-tune a model on all our contracts, so it knows them." What would you do instead?
To keep the contracts inside, run an open-weight model on your own computers. First check that its license allows your use. Do not fine-tune it on facts. Facts learned that way get blurry and out of date. You cannot trace them to a source, and they are hard to delete. Use retrieval over the contracts, with citations. Keep fine-tuning for a habit that a good prompt cannot fix. See Prompt, RAG or fine-tune?
Q5 · Priya changes a prompt. She tries 3 questions and loves the answers. On her 20-case test set, the pass rate goes up from 80% to 85%. But 2 "not in the sources" cases that passed before now fail. Should she ship it?
Not yet. The better average hides a regression in the cases that matter most. The new prompt now makes up answers it should refuse. Compare case by case. Fix the prompt, for example: "If the answer is not in the sources, say so." Run the whole set again. Ship only when nothing that passed before now fails. See Evals: does your AI work?
Go deeper: a checklist before you rely on AI
- Could the fact be newer than the model, or private? Then it must come with the question. See Search by meaning.
- Did it cite a source? Did I open it? See When AI makes things up.
- Did my question show the answer I hoped for? Then ask again, neutrally.
- Is my brief complete: goal, readers, format, examples and "say if you are not sure"? See Prompts that work.
- Is this the right size, effort and speed? Is it the right place for this data? See Choosing a model.
- Is this really a choice from a list? See Models that pick from a list.
- Is the problem missing knowledge, a wrong habit, or not an AI problem at all? See Prompt, RAG or fine-tune?
- Did I run the eval again after the change? See Evals.
That's AI essentials. You can explain what a model, a token and a context window are. You can tell when an assistant may be wrong, or may just agree with you. You can write a good brief and pick a model. You can choose the right fix, and prove that a change helped. Next up — Working with AI: APIs, tools, agents and automation. Start at the Working with AI start page.
Pick one task you do every week. Write a good brief for it. Test it on 5 real examples with any AI assistant before you trust it.
Want more? Browse our free security tools on the tools shelf.