Tokens, agents, RAG, MCP, fine-tuning. Here is what each one actually means - no maths, no hype.
Also called: agentic
A process where an AI model runs several steps on its own, with checkpoints where a human or another system can step in.
Why it mattersMost real business automation lands here - not full autonomy, not a single prompt.
See also AI Agent, Human in the Loop, n8n
Also called: agent, autonomous agent
Software that takes a goal, decides its own steps, and uses tools to get there - instead of following a fixed script. It loops: think, act, check the result, try again.
Why it mattersThis is the shift everyone's talking about. A chatbot answers you; an agent goes and does the thing.
See also Agentic Workflow, Tool Use, Orchestration
The work of making a model behave the way people actually intend, including when nobody wrote a rule for the situation.
Why it mattersIt's why two models with similar raw ability can feel very different to work with.
See also RLHF, Guardrails
Also called: application programming interface
A way for your code to talk to a service. You send a request, you get a structured response back.
Why it mattersAnything beyond copy-pasting into a chat window goes through an API.
See also API Key, Rate Limit
Also called: secret key
A private string that identifies your account when you call an API. Treat it like a password.
Why it mattersLeaked keys get scraped from public repos within minutes and run up real bills.
See also API
Also called: AGI
A hypothetical system that can handle any intellectual task a person can, rather than being good at a narrow set.
Why it mattersNothing shipping today is AGI. When a vendor implies otherwise, that's marketing.
See also Artificial Intelligence
Also called: AI
The broad field of getting machines to do things we'd call intelligent - recognising, predicting, reasoning, generating.
Why it mattersIt's an umbrella, not a product. Ask what's underneath when someone says a tool is "AI-powered".
See also Machine Learning, Generative AI
Also called: self-attention
The mechanism that lets a model weigh which earlier words matter most when producing the next one.
Why it mattersIt's the idea that made transformers work, and the reason models track context across long passages.
See also Transformer, Context Window
Also called: eval, leaderboard
A fixed set of tasks used to score models against each other. MMLU, SWE-bench, GPQA and friends.
Why it mattersUseful for rough ranking. A benchmark win rarely predicts whether a model is good at your job.
See also Evals
Skew in a model's output that traces back to skew in its training data or design choices.
Why it mattersIt shows up quietly - in who gets shortlisted, whose accent transcribes badly, which photo it picks.
See also Alignment, Explainability
Also called: CoT, step by step
Getting a model to write out its intermediate reasoning before answering, rather than jumping to a conclusion.
Why it mattersCheapest accuracy upgrade there is on multi-step problems. Just ask it to work through the steps.
See also Reasoning Model, Prompt Engineering
Also called: splitting, segmentation
Cutting documents into smaller passages before storing them for retrieval.
Why it mattersBad chunking is the number one reason a RAG system returns nonsense. Split on meaning, not character count.
Also called: coding agent, CLI agent
A coding agent that works in your terminal or editor - it reads your repo, edits files, runs commands, and opens pull requests.
Why it mattersDifferent from autocomplete. You hand it a task, not a line.
See also AI Agent, Vibe Coding, MCP
Also called: context length
The maximum amount of text a model can hold in view at once - your prompt, the conversation so far, and its reply, all counted in tokens.
Why it mattersRun past it and the earliest content silently drops. It's usually the reason a long chat "forgets".
See also Token, Prompt Caching
Machine learning using neural networks with many layers, trained on large datasets.
Why it mattersEvery model you've heard of in the last decade sits in this category.
See also Neural Network, Machine Learning
Also called: model distillation
Training a small model to copy the behaviour of a big one.
Why it mattersHow you get near-frontier quality at a fraction of the cost and latency.
See also Quantization, Open-Weights Model
Also called: vector, embeddings
A list of numbers representing a piece of text, image or audio, arranged so similar things sit close together.
Why it mattersThe trick behind semantic search - it finds meaning, not keyword matches.
See also Vector Database, Semantic Search
Also called: evaluation, test suite
Your own test set for checking whether a prompt, model or agent does the job on your data.
Why it mattersWithout evals you're guessing. This is what separates a demo from something you can ship.
See also Benchmark, Hallucination
Also called: interpretability
How well you can tell why a model produced a particular output.
Why it mattersRegulated industries will ask. "The model said so" isn't an answer in a lending or hiring decision.
Also called: examples, one-shot
Including a handful of worked examples in your prompt so the model matches the pattern.
Why it mattersFaster and cheaper than fine-tuning, and it fixes most format problems on its own.
See also Zero-Shot Prompting, Prompt Engineering
Also called: SFT, supervised fine-tuning
Further training a pre-trained model on your own examples so it picks up a specific style, format or domain.
Why it mattersReach for it when prompting plateaus - and after you've tried retrieval first.
See also Pre-training, RAG, LoRA
Also called: base model, frontier model
A large model trained on broad data that gets adapted to many downstream tasks.
Why it mattersYou rent these rather than build them. Training one costs more than most companies are worth.
See also Large Language Model, Pre-training
Also called: tool calling
Letting a model return a structured request to run one of your functions, with arguments filled in.
Why it mattersThe bridge between text generation and doing something real - booking, querying, sending.
See also Tool Use, Structured Output, MCP
Also called: GenAI
Models that produce new content - text, images, code, audio - instead of only classifying or predicting.
Why it mattersThe part of AI that got everyone's attention in 2022 and hasn't let go.
See also Large Language Model, Multimodal
Tying a model's answer to specific source material it was given, so claims trace back to something real.
Why it mattersAn ungrounded answer sounds identical to a grounded one. That's the whole problem.
See also RAG, Hallucination
Also called: safety filters
Checks around a model that block bad inputs or outputs - policy filters, schema validation, spend limits.
Why it mattersAgents especially need them. Anything that can act can act wrongly.
See also Human in the Loop, Prompt Injection
Also called: confabulation, made up
When a model states something false with complete confidence - an invented citation, a plausible wrong number.
Why it mattersIt's not a bug they'll patch out. Design your process to verify anything that matters.
Also called: HITL, approval step
A checkpoint where a person reviews or approves before the automation continues.
Why it mattersThe difference between an agent that drafts the email and one that sends it to 4,000 customers.
See also Guardrails, Agentic Workflow
Running a trained model to get an answer. The everyday act of using AI.
Why it mattersTraining is a one-off cost. Inference is the bill that arrives every month.
See also Latency, Pre-training
A prompt crafted to get a model to ignore its own rules.
Why it mattersAssume anything user-facing will be probed. Filter what comes out, not just what goes in.
See also Prompt Injection, Guardrails
Also called: training cutoff
The date after which a model has no built-in knowledge of world events.
Why it mattersAsk about last week and you'll get confident nonsense unless the model can search.
See also RAG, Hallucination
Also called: LLM
A model trained on enormous amounts of text to predict what comes next, which turns out to be enough to write, summarise, translate and reason.
Why it mattersGPT, Claude, Gemini, Llama, Qwen - all LLMs. Everything else is packaging around them.
See also Foundation Model, Token, Transformer
Also called: speed, time to first token
How long a model takes to respond - often measured as time to the first token, then tokens per second.
Why it mattersA model that's 5% smarter but three times slower loses in most real products.
Also called: low-rank adaptation, adapter, PEFT
A cheap fine-tuning method that trains a small set of extra weights instead of the whole model.
Why it mattersPuts custom models within reach of a laptop budget rather than a data centre one.
See also Fine-Tuning, Open-Weights Model
Also called: ML
Getting a system to learn patterns from data instead of being programmed with explicit rules.
Why it mattersThe engine under almost everything labelled AI, including the boring, useful stuff.
See also Deep Learning, Training Data
Also called: Model Context Protocol, connectors
An open standard for connecting models to tools and data sources, so one integration works across different apps and models.
Why it mattersStops every AI tool needing its own bespoke plug for your Gmail, Notion or database.
See also Tool Use, Function Calling
Also called: vision, image input
A model that handles more than one kind of input or output - text plus images, audio or video.
Why it mattersScreenshots, whiteboards, PDFs and invoices become usable input.
See also Generative AI
Also called: workflow automation, no-code
An open-source workflow tool where you wire together triggers, apps and AI steps visually, then self-host it.
Why it mattersThe fastest route from 'we should automate this' to something running, without writing a backend.
See also Orchestration, Agentic Workflow
Also called: neural net
Layers of simple maths units with adjustable weights, tuned by training until the whole thing maps inputs to useful outputs.
Why it mattersYou don't need the maths to use one - but knowing it's statistics, not magic, sets your expectations right.
See also Deep Learning, Parameters
Also called: open source LLM, local model
A model whose trained weights you can download and run yourself. Llama, Mistral, Qwen, DeepSeek.
Why it mattersYou control the data, the cost and the uptime. You also own the ops.
See also Quantization, Distillation
Also called: pipeline, chaining
Coordinating several models, tools and steps into one reliable process, with retries and error handling.
Why it mattersWhere most agent projects actually succeed or fail. The model is the easy part.
See also n8n, Agentic Workflow
Also called: weights, 7B, 70B
The learned numbers inside a model. '70B' means 70 billion of them.
Why it mattersMore parameters usually means more capable and more expensive - but training quality matters more than raw size.
See also Neural Network, Quantization
Also called: base training
The first, giant training run where a model learns general patterns from a broad dataset.
Why it mattersMonths of compute and a nine-figure bill. Everything after this is comparatively cheap.
See also Fine-Tuning, Foundation Model
Everything you send the model - instructions, context, examples, the actual question.
Why it mattersThe single biggest lever on output quality, and the only one that's free.
See also System Prompt, Prompt Engineering
Also called: caching
Reusing the processed form of a repeated chunk of prompt instead of paying to process it again.
Why it mattersLong system prompts and big documents get much cheaper and faster on repeat calls.
See also Token, Context Window
Writing and refining prompts systematically - clear instructions, examples, structure, constraints - until output is reliable.
Why it mattersLess mystical than it sounds. It's mostly being specific and testing.
See also Few-Shot Prompting, Chain of Thought, Evals
Also called: indirect injection
Hiding instructions inside content a model will read - a webpage, an email, a PDF - to hijack what it does next.
Why it mattersThe top security risk for agents. If your agent reads untrusted text, assume it can be talked into things.
See also Guardrails, Jailbreak
Also called: 4-bit, GGUF
Storing a model's numbers at lower precision so it needs less memory and runs faster.
Why it mattersHow a 70B model ends up running on a decent laptop with only a small quality hit.
See also Open-Weights Model, Distillation
Also called: retrieval augmented generation
Fetching relevant documents at question time and handing them to the model so it answers from your material.
Why it mattersThe standard way to make a general model useful on private, current company data.
Also called: throttling, quota
A cap on how many requests or tokens you can send in a window.
Why it mattersHits hardest at demo time. Build in retries with backoff before you need them.
Also called: thinking model, extended thinking
A model trained to spend extra compute working through a problem internally before answering.
Why it mattersWorth the extra seconds and cost on hard problems. Overkill for summarising an email.
See also Chain of Thought, Latency
Also called: reinforcement learning from human feedback
Training a model on human preferences between candidate answers, so it learns which responses people actually want.
Why it mattersThe step that turned raw text predictors into assistants worth talking to.
See also Alignment, Fine-Tuning
Also called: vector search
Searching by meaning rather than exact words, using embeddings.
Why it mattersFinds the doc about 'staff leave policy' when someone searched 'time off'.
See also Embedding, Vector Database
Sending the response token by token as it's generated instead of waiting for the whole thing.
Why it mattersSame total time, far better experience. Users watch progress rather than a spinner.
Also called: JSON mode, schema
Forcing a model to reply in a fixed shape - valid JSON matching a schema you define.
Why it mattersTurns text into something your code can use without brittle parsing.
See also Function Calling, Tool Use
Also called: system message
Standing instructions that sit above the conversation - role, tone, rules, what to refuse.
Why it mattersWhere you encode your product's behaviour. Users see the results, never the text.
See also Prompt, Prompt Injection
Also called: sampling, top-p
A setting for randomness. Low temperature gives consistent, predictable text; high gives varied, surprising text.
Why it mattersNear zero for extraction and classification. Higher for brainstorming and copy.
See also Inference
Also called: tokens, tokenization
The chunks a model reads and writes - roughly three-quarters of a word each in English.
Why it mattersYou're billed per token and limited by tokens. It's the unit of everything.
See also Context Window, Prompt Caching
Also called: tools, actions
Giving a model access to real capabilities - search, a database query, a calculator, an API call - that it can invoke when needed.
Why it mattersTurns a model that talks about your data into one that reads it.
See also Function Calling, MCP, AI Agent
Also called: dataset, corpus
The material a model learns from.
Why it mattersIts blind spots, biases and cutoff date all trace back here.
See also Pre-training, Bias, Knowledge Cutoff
The neural network architecture, introduced in 2017, behind nearly every modern language model.
Why it mattersThe 'T' in GPT. It's why progress accelerated so sharply.
See also Attention, Large Language Model
Also called: vector store, pgvector, Pinecone
A database built to store embeddings and find the nearest matches fast.
Why it mattersThe memory layer for retrieval. Start with pgvector before buying anything.
See also Embedding, RAG, Semantic Search
Building software by describing what you want to a coding agent and steering the result, rather than writing most lines yourself.
Why it mattersGreat for prototypes. Still needs review before anything touches production.
See also Claude Code, AI Agent
Asking a model to do a task with no examples, just instructions.
Why it mattersTry this first. Modern models handle far more zero-shot than people expect.
See also Few-Shot Prompting, Prompt
The 6-week cohort is where you build with them - small group, mentor-led, 20 seats.
See the cohort