The AI Minimalist Glossary. Everything you need to know about AI.
- Jonathan Chew
- Jun 12
- 6 min read
Updated: Jun 16
By Jonathan Chew, Founder of AI Minimalist
The only place you need to go for everything AI. Seven parts. One guide. Every term, tool, and player - in one place. No filler. No noise.
AI moves fast. The vocabulary moves with it. One week it is RAG and fine-tuning. The next it is MCP, agentic workflows, and vibe coding. If you are not keeping the terminology sharp, you are already behind the conversation.
This glossary covers seven distinct territories - from the foundational building blocks of how models work, to the tools, companies, and concepts shaping what comes next. It is a complete map of the landscape so you know where everything sits and where to go deeper.
"The goal is not to know everything. It is to know enough to ask the right questions - and to spot when someone else does not." - Jonathan Chew
Seven parts below. Read straight through or jump to what you need.
Part 01 - The Building Blocks
Foundational concepts.
The vocabulary every AI conversation rests on.
Every conversation about AI sits on top of these terms. Whether someone says the model is hallucinating, that they are using RAG to ground the outputs, or that they want to fine-tune for their use case - these are the words that frame everything. Knowing them well means you can evaluate claims, spot weak reasoning, and contribute something real.
Transformer - The architecture behind almost every major language model. Introduced in 2017. Still the foundation of GPT, Claude, Gemini, and most of what you use daily.
Tokens - The units models read and write in. Not words - chunks. Context windows are measured in tokens. Pricing is per token. "Unbelievable" might be 3 tokens.
RAG - Retrieval-Augmented Generation. Feeding a model relevant documents at inference time so it answers from real, current data - not just its training.
Hallucination - When a model produces confident-sounding output that is factually wrong. Not a bug in the traditional sense - a structural property of how these models generate text.
Fine-tuning - Further training a pre-trained model on a specific dataset to specialise its behaviour. Expensive but powerful for domain-specific tasks.
Context Window - How much text a model can hold in working memory at once. Modern models range from 8k to over 1 million tokens. Size shapes what tasks are even possible.
RLHF - Reinforcement Learning from Human Feedback. How models are trained to be helpful and follow instructions - humans rate outputs, those ratings shape the model.
Temperature - Controls output randomness. Low = predictable and precise. High = creative and variable. Most tasks sit somewhere in between.
Also in this section: Attention, Embeddings, Pre-training, Inference, Mixture of Experts, Reasoning Models, Multi-modal, System Prompt, Few-shot, Zero-shot, Foundation Model, Skills Engineering + more.
Part 02 - Pixels on Demand
Image generation.
The names doing the heavy lifting in pixels.
Image generation has gone from novelty to professional tool in under three years. The gap between models is real - different tools have different strengths in photorealism, illustration, text rendering, consistency, and creative control. Knowing the names means you can choose the right one for the job.
Midjourney remains the benchmark for artistic quality and stylistic range. Flux has become the go-to for open-weight flexibility. GPT-Image-1 brought strong text-in-image rendering. Firefly is Adobe's commercially safe option for professional workflows. Nano Banana is the newest entry pushing quality-to-speed ratios.
Tools in this section: Nano Banana, GPT-Image-1, DALL-E, Midjourney, Flux, Stable Diffusion, Ideogram, Recraft, Imagen, Firefly, Leonardo, Krea, Reve, Playground + others.
Part 03 - Frames per Prompt
Video generation.
The frontier moved this year. Know all the names.
Video generation was a curiosity in 2023. In 2025 it became a serious production tool. The release of Sora 2 and Veo 3 - which includes native audio - reset expectations across the industry. Consistency, physics, motion quality, and prompt adherence are all improving with each generation.
Runway established the professional standard early. Kling from China surprised the market with quality-to-cost performance. Higgsfield and Luma Dream Machine compete on cinematic realism. The category is crowded and moving fast - names that were leading six months ago may be displaced by the time you read this.
Tools in this section: Veo 3, Sora 2, Kling, Seedance, Runway, Luma Dream Machine, Pika, Hailuo, Wan, LTX, Hunyuan, Higgsfield, Genmo + others.
Part 04 - Voice & Sound
Audio generation.
Voice, music, and everything that vibrates.
Audio is split into two distinct tracks: voice and music. They have different players, different use cases, and different quality benchmarks.
ElevenLabs dominates professional voice cloning and text-to-speech, with near-human quality across dozens of languages. Suno and Udio lead music generation - full songs, any genre, from a text prompt. NotebookLM introduced a different format entirely - AI-generated podcast audio from your own documents. Cartesia focuses on ultra-low latency voice for real-time applications.
Tools in this section: ElevenLabs, Suno, Udio, Cartesia, NotebookLM, MusicLM, Hume, Resemble, Play.ht, Stable Audio, Riffusion, WaveNet + others.
Part 05 - Who's Who
The major players.
Labs, products, and the platforms shaping how you'll use AI.
The competitive landscape is more crowded than ever - and more concentrated. A handful of labs are pulling ahead in frontier model capability while dozens of well-funded challengers compete in specific verticals, open weights, and regional markets.
OpenAI - The company that triggered the current wave with ChatGPT in late 2022. Still at the frontier with GPT-4o, o3, and Codex. The most-used AI platform in the world by a significant margin.
Anthropic - Built by former OpenAI researchers, focused on AI safety and the Claude model family. Claude is widely regarded as the strongest model for long-context reasoning, coding, and written work.
Google DeepMind - Google's unified AI research arm. Produces Gemini, leads on multi-modal capability. Veo 3 is their video generation breakthrough. Has the largest compute infrastructure of any lab.
Meta & DeepSeek - Meta's Llama series and DeepSeek's R1 made high-quality open-weight models available to anyone. DeepSeek's efficiency breakthrough in early 2025 reshaped the cost conversation entirely.
Also in this section: Microsoft, xAI, Mistral, Cohere, Perplexity, Hugging Face, Alibaba Qwen, Amazon, Apple, NVIDIA, Inflection + others.
Part 06 - Build with AI
Coding tools.
The category that moved fastest. Everyone wants in.
No category attracted more investment than AI coding tools. From autocomplete to autonomous agents that open pull requests - the range of what these tools do is enormous.
Claude Code and Codex operate as fully agentic coding systems, capable of running terminals, writing tests, and iterating independently. Cursor became the editor of choice for developers who want AI deeply embedded in their workflow. Tools like Lovable, Bolt, and Replit have made vibe coding - building software through natural language - accessible to non-engineers.
The boundary between developer and builder is shifting faster than any job description has caught up with.
Tools in this section: Claude Code, Codex, Cursor, Antigravity, GitHub Copilot, Windsurf, vo, Bolt, Lovable, Replit, Aider, Continue, Tabnine, Devin, Junie, Zed, Cline, Augment + others.
Part 07 - Don't Get Caught Out
Concepts you should know.
The terms showing up everywhere now.
These are the terms that have graduated from technical jargon to everyday conversation. You will hear them in product meetings, conference talks, job descriptions, and funding announcements. Knowing what they actually mean - not just the surface definition - puts you ahead.
Agentic AI - AI systems that take sequences of actions autonomously - browsing the web, writing and running code, sending messages - without a human approving each step.
MCP - Model Context Protocol. Anthropic's open standard for connecting AI models to external tools and data sources. Rapidly becoming the integration layer of the agentic stack.
Vibe Coding - Building software through natural language prompts rather than writing code directly. Popularised by Andrej Karpathy. Now a real production workflow for many builders.
Chain-of-Thought - Where the model reasons step by step before giving an answer. Significantly improves performance on complex tasks.
Alignment - The challenge of making AI systems reliably do what humans intend - and avoid what they do not. One of the core research problems at every frontier lab.
Guardrails - Constraints placed on model behaviour - filters, rules, and trained boundaries that prevent outputs from going outside acceptable limits.
Distillation - Training a smaller model to replicate the behaviour of a larger one. How DeepSeek produced high-quality small models at a fraction of the typical cost.
Synthetic Data - Training data generated by AI models rather than collected from humans. Now central to how frontier models are trained - models learning from models.
Also in this section: Quantisation, Edge AI, Constitutional AI, Tool Use, Function Calling, Memory, Jailbreak, Governance + others.
Want the whole thing?
Seven parts. Every term. Every link.
Read the complete AI Minimalist Glossary on Medium - curated by Jonathan
Chew, Founder of AI Minimalist.
7 Parts - One Guide - No Noise
For Faster Learning
AI Minimalist Resource Vault
Over 300 hand-picked resources covering AI tools, frameworks, and practical workflows. Built to give you what you need to start applying these skills right away. No filler.
For Lasting Foundations
UBI x BrandRev.ai Scholarship
Resources help you move fast. But the professionals who stay ahead long-term are the ones who built real foundations first. We have partnered with UBI Business School to make a structured, internationally recognised AI qualification more accessible. Scholarships are available. If you want depth that holds beyond the next quarter, this is worth looking at.
Found this useful? Pass it on to one person who is spending more time chasing AI tools than actually using them well.
Comments