AI models trained on enormous amounts of text. They predict the next likely word. ChatGPT, Claude and Gemini are all LLMs.
On the list because: When people say "AI" in a product context, they almost always mean an LLM. You need to know the word.
A large, general AI model trained broadly that can be adapted to specific tasks. Claude, GPT, Gemini.
On the list because: Most companies build on top of a foundation model instead of training their own. That is a strategic dependency you need to be able to see.
How much text an AI model can hold in a single conversation. Measured in tokens.
On the list because: It sets the practical limits for what AI can do in your product. Bigger is not always better: models get worse at recalling details in very long context.
The unit AI models work in. One token is roughly 0.75 English words.
On the list because: It is how AI is billed. Understand tokens and you can estimate costs. Fail to understand them and the bill arrives as a surprise.
Text turned into numbers, a long list of coordinates that places the meaning of the text in a mathematical space.
On the list because: It is the engine behind intelligent search and the foundation of RAG. If your product needs to search your own documents meaningfully, you end up here.
An architecture where the AI fetches relevant documents from your own knowledge base and uses them to answer. The AI "looks things up" before it answers.
On the list because: It is the most widespread way to get AI to answer from your specific data. Most internal AI assistants are built on RAG in some form.
Anthropic’s open standard for how an AI model gets access to external data and tools, e.g. your CRM, Notion, Gmail.
On the list because: It is becoming the industry standard. If you pick a proprietary solution over MCP, you can end up locked to one vendor. A strategic choice, not just a technical one.
An AI system that can carry out tasks autonomously across multiple steps: decide what to do, use tools, check whether the result is right, and try again.
On the list because: It is the difference between AI as a chatbot and AI as a coworker. Most "AI features" being built right now are agents, even when marketing calls them something else.
The AI model’s ability to call external functions: look something up in a database, send an email, run a calculation.
On the list because: It is how an agent actually gets anything done. Without tool calling, AI is just text generation.
The discipline of phrasing instructions to AI so the model delivers what you want.
On the list because: It is a real competence, not an afternoon skill. Research shows it predicts output quality.
The underlying instruction that defines how the AI should behave: its role, rules, boundaries.
On the list because: It is the most important design choice when you build on top of an LLM. A good system prompt turns a generic model into a specific product.
Reusable packages of instructions, examples and tools that an AI agent loads on demand, a "skill" the agent picks up exactly when the task calls for it, instead of holding everything in its head at once.
On the list because: It is the way to make a general agent specialised without training a new model. Instead of one enormous system prompt, you split knowledge into skills the agent fetches when it needs them: cheaper to run, easier to maintain, and the content can be owned by the product organization rather than only the developers.
Systematic testing of whether an AI model does the right thing. Think: unit tests for probabilistic systems.
On the list because: You cannot debug an LLM like ordinary code. Evals are the precondition for improving AI features systematically instead of by gut feeling.
When AI generates something that looks credible but is wrong or made up.
On the list because: It is the single biggest threat to trust in AI products. People forgive obvious mistakes. They do not forgive mistakes that look right but are lies.
The practice of anchoring AI answers in specific, verifiable sources, e.g. your documents, a database, or real-time data.
On the list because: It is how you make AI usable in a serious product context. An AI that cites where it knows something from is a tool. An AI that just "knows things" is a risk.
Two ways to get AI to do something specific. Fine-tuning is training the model further on your data. Prompting is instructing it without changing the model.
On the list because: Most people start with fine-tuning because it sounds more "real". Best practice is the opposite: prompt until you have evidence that it is not enough.
The collective ability of the organization to understand, assess and use AI responsibly. The whole organization, not just the technical people.
On the list because: It is the underreported bottleneck in most "AI transformations". People cannot use tools they do not understand.
Building software by describing what you want and letting AI generate the code, without writing it line by line yourself.
On the list because: It is shifting who can build software. Product managers, designers and founders now build apps in a weekend. You need to know the term.
A design pattern where AI takes steps toward a task, but a human approves, corrects or finishes before anything happens externally.
On the list because: It is where most serious AI features should land over the next few years, especially in regulated industries.
The constraints you build into an AI system to prevent unwanted outputs.
On the list because: They are product decisions disguised as technical choices. Every guardrail is an answer to "what must our AI not do?", and the product organization should help answer that.
An AI model actually generating an answer, the moment where input goes in and output comes out.
On the list because: Inference is where AI costs money in production. Training is a one-off expense; inference is what you pay every time a user interacts with your product.
Missing a term? Write to me. I add entries continuously.