"AI Glossary: 40 Terms You Should Know (2026)"

The AI terms behind the tools, in plain English.

Foundations

• LLM — large language model; predicts the next token from patterns in training data.
• Token — a chunk of text (roughly 4 characters); models bill and limit by tokens.
• Transformer — the neural architecture behind modern LLMs.
• Multimodal — handles text, image, audio, and video together (e.g., GPT-4o).

Training & tuning

• Pre-training — learning from a huge corpus.
• Fine-tuning — further training on a specific dataset.
• RLHF — reinforcement learning from human feedback; aligns outputs with helpful/safe answers.
• Embedding — a vector representation of text, used in search and RAG.
• LoRA — a small fine-tune layered on a base model (common in Stable Diffusion).

Capabilities

• Hallucination — fluent but false output.
• Context window — how much text the model can ‘see’ at once (e.g., 1M tokens).
• RAG — retrieval-augmented generation; answers from your data.
• Agent — a system that plans and acts with tools toward a goal.
• Prompt — the instruction you give the model.

Image & video

• Diffusion — the process behind most image generators (gradual denoising).
• Text-to-image / video — generates from a prompt.
• Image-to-video — animates an uploaded image (Luma, Kling).
• Lip-sync — matching mouth movement to generated audio (Veo, HeyGen).

Voice & music

• TTS — text-to-speech (ElevenLabs, Murf).
• Voice cloning — replicating a specific voice from samples.
• Stems — isolated instrument tracks (Udio, Suno Premier).

Business

• AdSense — Google’s ad network for publishers.
• Commercial rights — permission to use output commercially (check each tool).
• API — programmatic access to a model.

FAQ

What is a token, practically? About 4 characters of English; ~750 words is ~1,000 tokens. Models cap input and bill by tokens.

What is the difference between RAG and fine-tuning? RAG feeds context at query time; fine-tuning changes the model. RAG is cheaper to update.

Why does context window matter? Bigger windows (1M tokens) let the model reason over whole codebases or books at once.

Verdict

You do not need a CS degree, but these 40 terms decode every tool review on OperatorStack. Bookmark it.