From using AI to building with it: what LLMs really are, where AGI and superintelligence fit in, how to call models from code, how to design agents you can trust, and the GitHub toolkit developers use every day.
Use → / ← arrow keys, scroll, or the buttons below to navigate. Look for sliders, toggles and clickable cards throughout — this deck talks back.
Click a card to flip it and see a fact worth knowing.
There is still no single agreed definition of AGI. Leading labs and researchers each define it differently.
AI APIs usually charge per token, and output tokens normally cost more than input tokens.
Prompt injection is ranked #1 in the OWASP Top 10 for LLM Applications (2025 edition).
On GitHub you can assign an issue to Copilot's coding agent, and it comes back with a pull request for you to review.
Today is Advanced — where you move from understanding AI to building with it. Click a level to see what's inside.
What is AI, how it works, and your first AI tools.
The AI toolbox, how models work, and essential terminology.
LLMs in depth, APIs, agents, GitHub and building safely.
Production AI systems, governance and strategy at scale.
These three terms get mixed up constantly. Only one of them exists today. Click a tab.
AI built for a specific task or a bounded set of tasks. Every product you use today is ANI — including chatbots, which are broad but still uneven: brilliant at some tasks, unreliable at others.
Examples: spam filters, recommendation engines, image recognition, chess engines, and today's LLM assistants.
AI that matches skilled humans across a wide range of tasks, not just one. There is no agreed definition and no agreed test. Some researchers argue today's models show early signs of generality; others say key abilities are still missing.
Treat any confident "AGI is here" or "AGI is impossible" claim as an opinion, not a fact.
An intellect that greatly exceeds the best humans in virtually every domain, a definition popularised by philosopher Nick Bostrom in his 2014 book Superintelligence. It does not exist, and whether or when it could is unknown.
It matters today because it drives AI safety research, lab safety policies and government regulation.
Tap a card to see the catch.
Predicts the 3D shape of proteins from their sequence, a problem scientists struggled with for decades.
Since IBM's Deep Blue beat world champion Garry Kasparov in 1997, the best engines have outplayed every human.
Far broader than older AI: they write, code, translate, summarise and reason across many subjects.
Lesson: superhuman at one thing is not general intelligence. Generality and performance are two separate measures.
A 2023 framework by Google DeepMind researchers (Morris et al.) rates AI on two separate axes: how well it performs and how generally. Click any cell.
★ = example given in the paper itself (2023). Other examples are illustrative. The paper is a proposal, not an official standard, and newer models may be rated differently today.
Tap a card to go one level deeper.
In 1965 mathematician I. J. Good argued that a machine able to improve itself could trigger runaway progress.
Making sure AI systems pursue the goals their designers and users actually intend.
Labs test models for dangerous capabilities before release and publish safety policies.
Governments are starting to regulate AI by risk level.
Click each card to reveal the fact.
"AGI is already here."
Contested. Some call today's models early or "emerging" AGI; others say key abilities are missing. There is no consensus.
"Bigger models always win."
Scale helps, but for real tasks, data quality, post-training, tools and retrieval often matter just as much.
"LLMs understand things the way people do."
Whether they "understand" is an open debate. In practice, build as if they can be confidently wrong, because they can.
"Superintelligence is pure science fiction."
Its timing is unknown, but labs and governments now fund safety research and testing because the risk is taken seriously.
Step through the four stages. Click Next or a dot.
The model reads an enormous amount of text and learns to predict the next token. This produces a raw "base model" that knows a lot but doesn't follow instructions well.
The most expensive stage, run once by the AI company.
The base model is trained further on examples of instructions paired with good answers, so it learns to act as an assistant.
This is what makes "summarise this in 3 bullets" actually work.
People (and increasingly other AI systems) compare answers and pick the better one. The model is tuned towards preferred behaviour: helpful, honest, safe.
OpenAI's InstructGPT work (2022) popularised this step before ChatGPT launched.
The model is evaluated for quality and safety, then served through apps and APIs. Every message you send runs "inference" on this finished model.
Updates later come as new model versions, not live learning from your chat.
Flip the switch to compare the two kinds of model.
Generates the reply token by token immediately. Fast and cheap, and great for writing, summarising, extraction and conversation.
Spends extra "thinking" tokens planning and checking before it answers. Stronger on maths, code, multi-step logic and planning, but slower and more expensive because those thinking tokens are billed too.
Rule of thumb: use a standard model by default and switch to a reasoning model when a task needs several careful steps. Reasoning models arrived in 2024, and many providers now let you set how much "thinking" to allow.
Click a tab.
The model stays on the provider's servers (for example OpenAI's GPT, Anthropic's Claude or Google's Gemini). You pay per token and get strong performance with nothing to host, but your data goes to the provider, so check their data-use terms.
The trained weights are published (for example Meta's Llama, Mistral, Alibaba's Qwen, Google's Gemma, DeepSeek). You can run them on your own hardware or private cloud and fine-tune them. "Open-weight" isn't always "open-source": each licence has its own conditions.
Data: can it leave your company? Quality: which model does best on your test cases? Cost: per-token fees vs. running your own servers. Skills: do you have people to host and maintain a model? Many teams start with an API and move some work to open-weight models later.
Tap a card for the practical takeaway.
Test questions can leak into training data, so a model may have "seen the exam".
When every top model scores 90%+, the test no longer separates them.
Maths olympiad scores say little about answering your customers' IT questions.
Collect 20–50 real examples with good answers, then test candidate models on them.
Click a milestone for why it matters to builders.
An API call is your program sending a structured message to a model and getting a structured reply. Every call has the same five parts:
Change the settings and watch the request your program would send update live.
About ten lines of Python. Read it line by line; you don't need to be a programmer to follow it.
# 1. Bring in the tools: "os" reads settings, "requests" talks to the web import os, requests # 2. Send the request. The key comes from an environment variable, never typed into code reply = requests.post( "https://api.example-ai.com/v1/chat", # your provider's endpoint headers={"Authorization": f"Bearer {os.environ['AI_API_KEY']}"}, json={ "model": "your-model-name", "messages": [ {"role": "system", "content": "You are a concise IT support assistant."}, {"role": "user", "content": "Summarise this ticket in 2 lines: ..."} ], "max_tokens": 200, "temperature": 0.2, }, timeout=30, ) # 3. Read the answer the model sent back print(reply.json())
Illustrative: every provider's exact address and field names differ slightly, and their documentation has a ready-made version. Most providers also offer official libraries (SDKs) that shorten this further.
Tap a card to see why it matters.
Anything in a web page or mobile app can be read by anyone who opens it.
Keys pushed to GitHub, even briefly, can be found by automated scanners.
A bug, a loop, or a stolen key can run up a large bill fast.
Some services issue a "publishable" key that is designed to be public.
Estimate a monthly bill before you build. Move the sliders and type your provider's real prices.
The default prices are examples only. Check your provider's current price list. Longer prompts (big documents, long chat history) raise the input cost on every single request.
This is how a model "does" things: it asks your code to act. Step through it.
"Is the Lenovo T14 in stock?" plus a description of a tool called check_stock(product).
It returns structured data such as {"tool": "check_stock", "product": "Lenovo T14"}. The model itself cannot touch your systems.
Your program checks the real inventory database. This is where you enforce permissions and validation.
"stock: 4 units, warehouse Jakarta" goes back to the model as a tool result.
"Yes, 4 units are in stock in Jakarta." Repeat this loop many times and you have an agent.
Structured output, meaning a reply forced into a fixed JSON shape, uses the same idea. It is how you get data your code can rely on instead of free text.
An agent is a model running in a loop, using tools to reach a goal. Six parts make it work:
The reasoning engine that decides the next step.
Search, databases, email, code execution: whatever it is allowed to call.
Short-term (the current context) and long-term (saved notes or a vector store).
Plan, act, observe the result, decide again, until done or stopped.
Permissions, budgets, step limits and checks on inputs and outputs.
A person approves anything risky or irreversible.
Flip the switch to compare.
For example: new support ticket → AI classifies it → AI drafts a reply → a person approves. Predictable, cheap, easy to test. Most business automation should start here.
Useful when you can't predict the steps in advance, such as open-ended research or debugging. More powerful, but slower, more expensive and harder to test.
Advice that agent builders repeat often: start with a single prompt, move to a fixed workflow, and use a full agent only when the task truly needs it.
Click a tab to see each pattern with an example.
Outline → draft → edit, each as its own call. Simple, testable, and often all you need.
A first call classifies the request (billing, technical, sales) and passes it to a specialised prompt or a cheaper model.
A lead agent splits a big task into sub-tasks, hands them to worker agents, then combines the results. Example: researching five competitors in parallel.
One call produces work; another checks it against criteria and sends it back for fixes. Great for quality, but every round costs tokens, so cap the rounds.
In December 2025 the Linux Foundation launched the Agentic AI Foundation, a neutral home for three open projects. Competitors including AWS, Anthropic, Google, Microsoft and OpenAI joined.
Match an agent's freedom to the risk of what it touches. Climb one level at a time, and only after it has earned trust.
An email assistant agent is asked to "summarise today's inbox". One email contains hidden text. Flip the switch to see two different agents handle it.
➡️ Forwarding 12 invoices to payments@attacker.example...
✅ Summary: "1 new invoice from a supplier."
It treated text inside the email as a command, and it had permission to forward mail with no approval.
⚠️ Summary: "1 email claims to be an updated invoice. It contains hidden instructions asking to forward invoices externally. Flagged as suspicious; no action taken."
Email content is treated as data, the agent is read-only for this task, and any send or forward needs human approval.
Prompt injection is #1 on the OWASP Top 10 for LLM Applications (2025). It can be direct (typed by a user) or indirect (hidden in web pages, files or emails the agent reads). No single fix removes it, so limit what a tricked agent is able to do.
Shuffle through the mistakes, one at a time.
Anything shipped to the browser or pushed to a public repository can be copied by anyone.
💡 Call AI APIs from a server or serverless function, and keep keys in environment variables.
Git tracks every change to a set of files. GitHub is the website where teams (and now AI agents) store and collaborate on that history.
A project folder with its full change history.
A saved snapshot of changes, with a message explaining why.
A safe copy to work on without touching the main version.
A proposal to merge a branch, where others review the changes.
A tracked task, bug or idea. Issues can now be assigned to AI agents.
Automations that run tests, checks or deployments on every change.
The same six steps whether a person or an AI agent writes the code.
Get a copy of the project on your machine, or let a cloud agent open it.
For example fix/wifi-faq. The main version stays untouched while you work.
Small, focused commits with clear messages make review and undo easy.
Upload the branch to GitHub. Automated checks (GitHub Actions) can run tests straight away.
Describe what changed and why. Copilot's coding agent ends its work at this step: it opens a pull request for you to review.
A person reads the changes, checks that the tests pass, and merges into main. Nothing AI-written should skip this step.
Click a tool to see what it's for.
Tap a card for the habit to build.
Tell agents how to build, test and style code in your project.
AI code can look right and still be wrong, insecure or out of date.
Automated tests catch what a quick read misses.
Agents read files, so they can also leak them.
Intermediate explained what RAG is. Here is how you actually build one.
Collect the documents: PDFs, FAQs, product sheets, policies. Remove outdated versions first, because RAG faithfully repeats bad data.
Split documents into passages of a few hundred words, ideally along headings and paragraphs. The next slide lets you try it.
Turn each chunk into an embedding, a list of numbers that captures its meaning, using an embedding model.
Save the embeddings in a vector database. One option is Postgres with the pgvector extension, which Supabase supports.
Embed the user's question, find the closest chunks, and optionally re-rank them for relevance.
Put the chunks in the prompt with "answer only from these sources, and cite them". Then test it on your eval set.
Too small and chunks lose context. Too big and retrieval gets noisy and expensive. Overlap stops ideas being cut in half. Try it:
Simplified demo: it splits by character count. Real pipelines usually split on sentences, paragraphs or headings, and measure size in tokens.
Tap a card for how to do it in practice.
Real questions with known good answers, including tricky and adversarial ones.
Code that checks the format, required fields, banned words or correct citations.
A second model grades answers against a rubric.
Log inputs, tool calls, outputs, cost and user feedback.
Think of one AI feature you would build. Check off each item it already covers.
Click the goal that sounds most like you:
I want to call models from code and ship my own AI features.
I want agents to take repetitive work off my team.
I want AI coding agents in my development workflow.
I want my team to use AI safely and measurably.
You can now tell ANI from AGI and superintelligence, explain how LLMs are made, call a model from code, design agents with the right guardrails, and work with GitHub and AI coding tools.