AI Infographics

Core AI concepts, drawn on demand

← AI Kaleido

Infographic Gallery

Hover to pause · tap a card to enlarge

Every picture is generated fresh, so give it a few seconds to appear.

Neural Networks

Like a huge relay team of tiny switches, each passing on a slightly refined guess.

  • Layers of neurons take numbers in and pass numbers out
  • Weights decide how strongly signals travel
  • Stacking layers lets simple parts learn complex patterns

Training vs Inference

Training is studying for the exam; inference is sitting the exam.

  • Training adjusts weights using labelled examples and errors
  • Inference runs the frozen model on new input
  • Training is slow and costly, inference is fast and repeated

Transformers & Attention

Every word looks around the room and decides who to listen to.

  • Attention scores link each token to the tokens that matter
  • Parallel processing replaced word-by-word reading
  • Foundation of nearly every modern language model

Tokens & Context Window

The model reads in Lego bricks, and its desk only fits so many bricks.

  • Text is split into tokens, roughly word fragments
  • The context window is the maximum tokens it can see at once
  • Older tokens fall out of view in long conversations

Embeddings

Meaning turned into map coordinates — similar ideas sit close together.

  • Text becomes a vector of numbers
  • Distance between vectors approximates similarity in meaning
  • Powers search, recommendations and retrieval

RAG — Retrieval Augmented Generation

An open-book exam: fetch the right page first, then answer.

  • Question is embedded and matched against a knowledge base
  • Retrieved chunks are pasted into the prompt
  • Reduces hallucination and adds fresh, private knowledge

Fine-tuning vs Prompting

Prompting is giving instructions; fine-tuning is retraining the employee.

  • Prompting shapes behaviour at request time, no training
  • Fine-tuning updates weights on your own examples
  • Fine-tune for style and format, prompt or retrieve for facts

AI Agents & Tools

A model with hands: it can look things up, click and act, not just talk.

  • Agent plans, calls a tool, reads the result, repeats
  • Tools include search, code, databases and APIs
  • The loop stops when the goal is reached

Hallucination

A confident student guessing an answer that sounds right but isn't.

  • Models predict likely text, not verified truth
  • Gaps in knowledge get filled with plausible invention
  • Grounding, retrieval and citations reduce it

Generative AI vs Traditional AI

Traditional AI sorts the mail; generative AI writes the letter.

  • Traditional models classify or predict a label
  • Generative models create new text, images, audio or code
  • Both learn from data, they just output different things

Guardrails & Alignment

Bumpers on a bowling lane — the ball still rolls, just not into the gutter.

  • Alignment tunes models toward helpful, safe behaviour
  • Guardrails filter inputs and outputs at runtime
  • Human feedback is used to rank preferred answers

The AI Project Lifecycle

Cooking: source ingredients, prep, cook, taste, serve, keep tasting.

  • Data collection and cleaning dominate the effort
  • Model build and evaluation come next
  • Deployment needs monitoring and retraining