AI Cheatsheet
How to become the "AI expert" during your family dinner
Artificial intelligence shapes our daily lives, from the recommendations on Netflix to the voice assistant on your phone. Yet the terminology can feel impenetrable. This comprehensive cheatsheet changes that, providing clear definitions of essential AI terms with practical examples and connections that help you understand how everything fits together.
Unlike technical manuals that assume expertise, this guide starts with fundamentals and builds systematically. Each term includes real-world applications and connects to related concepts, creating a complete picture of how AI actually works.
Foundation Concepts: Where It All Begins
Artificial Intelligence (AI): Computer systems capable of performing tasks that typically require human intelligence, learning, reasoning, problem-solving, perception, and decision-making.
Everyday reality: When your smartphone camera automatically focuses on faces, when Netflix suggests your next series, when your email filters spam—that's AI working behind the scenes.
Why this matters: AI isn't one technology but a family of approaches. Understanding this prevents both unrealistic fears and inflated expectations about what's possible today.
Algorithm: A set of step-by-step instructions that tells a computer how to solve a problem or complete a task.
As a simple analogy, we can think of it as a recipe for baking bread, specific instructions that, when followed precisely, produce a predictable result (nice smelling bread in this case).
In our daily lives, Google's search algorithm determines which websites appear first when you search. It follows millions of rules about relevance, authority, and user behaviour.
Every AI system relies on algorithms, but not all algorithms are intelligent. Your calculator follows algorithms but isn't classified as AI.
Machine Learning (ML): A subset of AI where computers learn to improve their performance by studying examples, rather than being explicitly programmed for every possible situation.
For example, unlike the early internet's way of programming "if email contains 'urgent money transfer', mark as spam", machine learning systems examine millions of emails and discover spam patterns themselves.
Machine learning provides the "learning" capability that makes artificial intelligence possible—it's how computers get smarter over time.
Learning Paradigms: How AI Systems Learn
Supervised Learning: Teaching AI by showing it examples with correct answers, like flashcards with questions on one side and answers on the other.
Medical example: Training an AI to detect cancer by showing it thousands of X-rays labelled "cancer" or "healthy" by expert radiologists.
Business application: Email spam detection, fraud prevention, speech recognition.
Unsupervised Learning: AI discovers hidden patterns in data without being told what to look for, like finding customer groups based on purchasing behaviour.
Marketing insight: Netflix uses unsupervised learning to group viewers with similar tastes, even when those similarities aren't evident to humans.
Key difference: Unlike supervised learning, no "correct answers" are provided—the AI finds patterns independently.
Reinforcement Learning: In this paradigm, AI learns through trial and error, receiving rewards for good decisions and penalties for poor ones.
Gaming breakthrough: How AlphaGo learned to beat world champions at Go—by playing millions of games against itself and learning from wins and losses.
We can think of self-driving cars learning to navigate by practising in simulators or autonomous trading systems optimising investment strategies.
Neural Networks: The Brain-Inspired Foundation
Neural Network: A computing system loosely modelled on the human brain, with interconnected processing units that work together to solve problems. Like neurons in your brain, artificial neurons receive signals, process them, and pass information to other neurons in the network. This technology enables pattern recognition that traditional programming can't achieve such as recognising faces, understanding speech, translating languages.
Deep Learning: Machine learning using neural networks with multiple layers (typically three or more) that can learn increasingly complex patterns (i.e., layer logic)
Early layers might detect edges in images, middle layers recognise shapes, and deeper layers identify objects like cats or cars.
This technology powers image recognition, natural language processing, autonomous vehicles, and medical diagnosis systems.
"Deep" refers to the multiple layers that make these networks powerful enough to handle complex real-world problems, which bridges this concept with neural networks.
Advanced AI Systems: Today's Cutting Edge
Large Language Models (LLMs): AI systems trained on vast amounts of text that can understand and generate human-like language across diverse topics.
Prominent examples include ChatGPT, Claude, and Grok, which are systems that can write, analyse, code, and converse naturally.
LLMs are built using transformer architecture and deep learning, representing the current pinnacle of natural language AI.
Despite human-like responses, these systems don't truly "understand" meaning; instead, they predict likely word sequences based on training patterns.
Transformer Model: The neural network architecture that revolutionised AI by using "attention mechanisms" to understand relationships in sequential data like text.
2017 breakthrough: This architecture made modern LLMs possible by dramatically improving how AI processes language and context. Unlike what many believe, Google is the real inventor behind the technology that enabled products like ChatGPT (pls refer to the original paper).
Attention innovation allows the model to focus on relevant parts of input when making predictions, similar to how humans focus on important words when reading.
The same technology powers not just language models but also image generation, code completion, and multimodal AI systems.
Generative AI: AI systems that create new content—text, images, music, code, or video—rather than just analysing existing content.
Creative applications include DALL-E, generating artwork from descriptions, GitHub Copilot writing code, and AI composers creating original music, which has been transforming industries from advertising and journalism to software development and entertainment.
Generative AI products often use generative models trained on massive datasets to learn patterns and create novel content.
Modern AI Capabilities and Applications
Natural Language Processing (NLP): AI's ability to understand, interpret, and generate human language in meaningful ways.
Daily examples include Apple's Siri understanding your questions (just joking!), Google Translate converting languages, and Gmail's smart compose suggesting email responses.
Recently, this technology has advanced from simple keyword matching to understanding context, nuance, and generating human-quality text.
Currently, large language models represent the current state-of-the-art in NLP, demonstrating unprecedented language capabilities.
Computer Vision: AI's ability to interpret and understand visual information from images and videos. A number of ubiquitous applications include face unlock on iPhones (not joking this time), automated checkout systems, medical image analysis, and quality control in manufacturing.
This technology is essential especially for self-driving cars, robotic systems, and augmented reality applications.
Multimodal AI: AI systems that can process and understand multiple types of data simultaneously—text, images, audio, and video.
Prompt Engineering: The skill of crafting effective (and efficient) instructions and questions to get optimal responses from AI systems, especially language models. Many people believed this would be the job of the future, until prompt engineering showed up to the party.
Technical Infrastructure and Considerations
Training Data: The information used to teach AI systems how to perform tasks, like textbooks for students, but on a massive scale.
This is by itself a separate topic, which can be highly technical. Still, if you use "scaling laws" in a sentence with a glass of red wine in hand, people are likely to respect you more.
The highly controversial nature of this topic for AI labs is apparent when we look at the sheer amount of incoming lawsuits, such as:
The New York Times has sued OpenAI and Microsoft in U.S. federal court, claiming the companies copied "millions" of pay-walled articles to train GPT models and now market products that can deliver those texts verbatim, threatening the paper's subscription business.
A consolidated class action led by the Authors Guild and 17 well-known fiction writers accuses OpenAI of wholesale ingestion of entire books—allegedly from pirate "shadow libraries"—and of stripping copyright-management information, seeking statutory and punitive damages under the DMCA.
Universal Music, Sony Music and Warner Records claim the Claude family of models was trained on unlicensed lyric databases. However, a judge recently denied a preliminary injunction; the labels continue to press copyright claims against Anthropic in Tennessee federal court.
Model Parameters: Internal settings that AI systems adjust during learning to improve performance, like tuning millions of dials to optimise output.
Interesting fact: GPT-3 has 175 billion parameters, while newer models exceed one trillion parameters.
Edge AI: The concept of running AI directly on local devices rather than in remote cloud servers. Privacy, privacy, privacy...
Overfitting: When an AI model learns training data too specifically, including noise and errors, making it perform poorly on new, unseen data.
We can think of this like a student memorising answers for specific test questions rather than understanding underlying concepts—fails when facing new questions. In order to prevent this, AI labs use validation datasets, regularisation techniques, and careful model design to ensure generalisation.
AI Safety and Ethics: Critical Considerations
AI Bias: Systematic prejudice in AI systems that can result in unfair or discriminatory outcomes, often reflecting biases in training data or design decisions. Such challenges generally originate from biased training data, incomplete datasets, or algorithmic design choices that reflect historical inequalities.
Explainable AI (XAI): AI systems designed to provide clear explanations for their decisions and reasoning processes. Personally, I like the "Interpretability" term better. If you struggle going to sleep, you can read this paper.
AI Safety: Research and practices focused on ensuring AI systems behave as intended and don't cause unintended harm. This is a quite comprehensive and fast-evolving topic, which I am also working on. To prevent this post from becoming a boring 25-page text, I kept it (very) brief.
Future Frontiers: Where AI Is Heading
Artificial General Intelligence (AGI): Hypothetical AI that matches human cognitive abilities across all domains—the ability to learn any task a human can learn. Clear definition, timing, and effects of AGI are still fluid. I previously posted a couple of pieces on this matter in case you are interested.
AI Governance: Frameworks, regulations, and standards for managing AI development, deployment, and impact on society.
Current initiatives include the EU AI Act, the UK AI Safety Institute, various national AI strategies and international cooperation efforts.
Btw, USA is currently considering a pause on regulating AI (shhh)
Let's Connect the Dots Together
The AI landscape forms an interconnected ecosystem where concepts build upon each other in a logical hierarchy. At the foundation, algorithms provide the basic instructions that enable Machine Learning to function. Machine Learning then encompasses Deep Learning as a specialised subset, which in turn powers Large Language Models and other advanced AI systems.
This technological stack connects to different learning approaches that span a broad spectrum. Supervised, Unsupervised, and Reinforcement Learning represent distinct methods for training AI systems, each requiring specific types of Training Data and proving suitable for different applications. These approaches work together rather than in isolation, often combining within single projects.
The development process follows a clear pipeline where Training Data feeds into these various learning approaches. This data then trains Neural Networks through careful management of Model Parameters whilst avoiding the common pitfall of Overfitting. Throughout this process, proper testing procedures validate that the system performs as intended.
These technical foundations support major application domains where AI makes its real-world impact. Natural Language Processing handles text and speech, whilst Computer Vision interprets images and video. Increasingly, these domains merge in Multimodal AI systems that can process multiple types of information simultaneously, creating more sophisticated and versatile AI applications.
Each component in this ecosystem depends on others, creating a web of relationships where understanding one concept helps illuminate the rest. This interconnected nature means that advances in one area often cascade through the entire system, driving progress across the field.