Bonus - AI glossary
AI glossary in plain English
Filter by category or type a term. Every entry has a short (one-line) and long version.
BasicsToolAn operation an agent can perform.+
Examples include reading a file, calculating a total or preparing a draft. An available tool is not permission to do anything with it.
BusinessApp connection (connector)Access to a service within granted permissions.+
For example, a drive connection provides access to permitted documents. Check the account and permissions. A skill describes how to work; a connection determines which sources can be reached.
BasicsAI (Artificial Intelligence)Systems performing tasks that require human intelligence.+
A broad field including machine learning and content generation. In business today "AI" usually means generative AI, i.e. large language models (LLMs) like ChatGPT, Claude, Gemini.
ModelsGPT-6 (Astra, Sol, Luna)OpenAI's newest model family (September 2026).+
Astra (3 Sep 2026) - the most capable, for hard tasks. Sol and Luna (22 Sep 2026) - cheaper models with near-Astra quality: Sol for everyday work and agents, Luna for fast tasks at scale. Choose a model by testing quality, time and usage on your own process. Check access and limits in your account; API billing and limits are separate. See lesson 13 for a worked cost example.
BasicsUsage capThe usage available in a plan.+
Check the limit and reset time in your account. Rules depend on the product, plan and model; do not assume team-wide pooling.
DevAPI billingSeparate billing for model and tool usage.+
API has its own prices and limits. Measure tokens, tool calls and iterations. A prompt instruction does not replace application-level cost controls.
ModelsLLM (Large Language Model)The engine behind ChatGPT and similar tools.+
A model trained on billions of texts to predict the next word. The GPT-6 (Astra, Sol, Luna) and GPT-5.6 (Sol, Terra, Luna) families, Claude, Gemini - all LLMs. "Large" = hundreds of billions of parameters.
BasicsTokenA unit of text processed by a model.+
A token can be a word fragment, character or whole word. Counts depend on the language, text and tokenizer; a word-to-token ratio is only an estimate.
BasicsPromptThe instruction you write for the AI.+
Prompt engineering means giving the model: role, context, goal, format and constraints. A good prompt helps, but the result still needs review.
ModelsContext / Context windowContext is information used to produce an answer; the window limits how much fits.+
Think of a workbench holding instructions, conversation and sources. It is not an archive or a guarantee that every sentence was used. For a large file, identify the relevant section and check what made it into the answer.
SecurityHallucinationWhen the model confidently states something made up.+
The model must always predict the next word - even without a real answer. Cure: turn on web search, ask for sources, verify facts.
ModelsTemperatureA sampling parameter, where supported by the model.+
Lower temperature does not guarantee truth. Availability and supported values depend on the model. Ask for sources and verify results regardless of this setting.
DevRAG (Retrieval Augmented Generation)AI + your knowledge base.+
Instead of retraining the model you plug it into your documents. It retrieves chunks and answers from them. That's how internal company chatbots work.
DevFine-tuningAdditional model training on examples.+
First compare prompting, retrieval and evaluation results. Fine-tuning can help with repeatable formats or behaviour, but needs suitable data and quality evaluation.
DevEmbeddingText turned into numbers (a vector).+
Lets you measure similarity between texts. The basis of RAG and semantic search - you search by meaning, not keywords.
ModelsAgentAn AI that runs multi-step tasks on its own.+
Doesn't just answer - plans, uses tools (calendar, mail, browser), checks results, iterates. Work mode in ChatGPT is an example of an agent for working with company files and apps.
DevMCP (Model Context Protocol)A standard for connecting AI to external tools.+
An open protocol (from Anthropic) - lets you plug any tool into the agent: Slack, Notion, databases, your own APIs. ChatGPT Work and Claude already use it.
DevAPIAn interface through which programs exchange data and request actions.+
For example, an app requests an order status and receives structured data. APIs are not exclusive to AI. Access may require authentication, and calls can have separate costs.
ModelsGPT (Generative Pre-trained Transformer)The architecture behind OpenAI models.+
"Custom GPT" in ChatGPT is different - your own assistant with files and instructions. Note: OpenAI is retiring custom GPTs (creating new ones is planned to end on 26 Oct 2026, existing ones run until 11 Dec 2026). Their successor is the plugin - see lesson 09.
ModelsMultimodalityAI understands text, image, voice, video.+
Today's models see photos, hear voice and generate images. In practice: paste an error screenshot - it explains. Paste an invoice - it turns it into a table.
BusinessCustom InstructionsPersistent instructions for ChatGPT (Settings → Personalization).+
You write once: "I'm an accountant at a Polish manufacturing company, answer concisely, in English, with numbers." Applies to all conversations. Cheapest hack for better answers.
BusinessMemoryChatGPT remembers facts about you across chats.+
You can say "remember that our product is called X." Turn it off in settings. WARNING: don't put sensitive data into memory.
BusinessProjectsFolders with files + instructions for repeatable work.+
You upload documents (offer, terms, customer base), set an answer style - ChatGPT remembers everything within that project. Great for teams.
BusinessWorkspaceAn organisation’s working space with members and settings.+
Workspace describes where and with whom you work. Business and Enterprise are plan names, not synonyms for the space. Check the account, membership and access before working. See lesson 03 for plans.
SecurityDPA (Data Processing Agreement)A GDPR data-processing agreement.+
Required by GDPR with your AI vendor. OpenAI provides a DPA for Business / Enterprise / API. Without a DPA you can't process personal data.
SecurityHallucination vs. factA plausible sentence is not evidence.+
Separate what the source supports from interpretation. A higher sales value does not prove a campaign worked. Request the source, open it and check the claim, date and units.
DevInferenceEach model call (each answer).+
More inferences = more cost. In business we measure cost per process ("1 handled ticket = 0.05 EUR"), not per person.
ModelsOpen-source LLMPublicly available models (Llama, Mistral, Qwen).+
You can run them on your own server - data never leaves the company. Requires IT and GPUs. Good for sensitive industries (healthcare, legal, defense).
BusinessPlugin (package)A package of instructions and integrations for tasks.+
A plugin may include skills, apps and reference material. Installing a package does not grant mailbox access. Check the connection and permissions separately. Example: a reporting package, spreadsheet connection and report-writing skill — lesson 09.
BusinessSkill (ChatGPT Work)Saved instructions for a repeatable task.+
A skill describes inputs, rules, outputs and checks. It grants no permissions and does not guarantee identical responses. Test it on two known cases and on missing data.
BasicsChat, Work and CodexThree modes of the same app.+
Chat is a conversation. Work carries out a whole task with your files and tools and hands back a finished file to review. Codex is the same kind of agent with a developer view: commands, code changes (a diff) and tests. Work and Codex share account limits. Lessons 02, 06, 07.
ModelsEffort (reasoning level)How long the model thinks before answering.+
Low = fast and cheap. High/Max = more careful, slower, uses more limit. Lesson 13.
BusinessShared limit poolOne limit shared by Work, Codex and agents.+
A long Work task reduces what is left for Codex. Check Settings › Usage. Lesson 03.
BasicsPluginA package that connects ChatGPT to an app and adds ready skills.+
E.g. the Gmail plugin lets it read mail and write drafts. Connect once in Settings › Plugins. Lessons 01, 09.
BasicsSkillA saved instruction you don't have to rewrite.+
Like a recipe: describe once how to write an offer, then call it with /offer. Lessons 01, 12.
BasicsComputer useAI clicks and types in programs on your computer.+
Codex can open Chrome or Excel and do steps like a person. Start with tasks that send and pay nothing. Lesson 11.
BusinessScheduled taskA command AI runs on its own at a fixed time.+
E.g. a daily 8:00 inbox summary. Lesson 12.
DevRepositoryA project folder with its change history, usually on GitHub.+
The analogy has a limit: an ordinary folder holds only the current files, while a repository remembers every saved version. Codex works inside a repository: reads, edits and shows what it changed. Lesson 07.
DevDiffA view of differences: what was removed and what was added.+
Like “track changes” in a word processor, but for project files. A line with a minus was removed, a line with a plus was added. It is not a quality score - only reading both lines shows whether the change is good. Lesson 07.
DevPull requestA bundle of changes submitted for review.+
It contains the changes, a description and test results. Merging approves it and adds it to the main version of the project; a person makes that decision. Lesson 07.
SecurityDraftA version to check before anything goes out.+
AI prepares the draft, a human approves and sends. Lesson 18.
BasicsSitesA small website built by ChatGPT.+
E.g. an event page or an internal form. More › Sites. Lesson 10.
ModelsReasoning modelA model that plans before it answers.+
Better at numbers, analysis and multi-step work. All GPT-6 models reason. Lesson 13.
SecurityApprovalThe moment AI asks for your permission.+
Before an agent sends, deletes or buys, it shows the plan and waits for your yes. Lesson 18.
Missing a term? Reach out - we'll add it.
More detailHow a model produces an answerTokens, context and hallucinations explained. Text estimates are illustrative and should not be used for billing.
Bonus - How AI works
How AI actually works
These six concepts explain conversation length, token usage and errors in answers.
The model generates next tokens
A model produces an answer by generating successive text units called tokens. It uses learned patterns and the available context. A coherent answer does not establish that its facts are true.
Tokens - the currency of AI
A token is a text unit, such as part of a word or a character. Estimates such as ~4 characters in Polish or ~0.75 words in English depend on the text and tokenizer. Use actual usage reports for API billing.
Context window
The context window limits the material available to the model while generating an answer. The example of 128k tokens (~250 A4 pages) is illustrative, not every model’s limit. For long material, check which parts were used; context is different from account memory.
Data cutoff
A knowledge cutoff describes the training knowledge range. Later information can come from supplied files or tools such as search. Check the source and date for current information.
Temperature and creativity
Temperature affects generation randomness if the model supports the parameter. Values of 0 and 1 may produce different variation, but they do not measure truth or act as a simple creativity switch. A style instruction does not change this parameter.
Hallucinations
A hallucination is an answer that presents invented or incorrect information as fact. It can sound convincing. Ask for a source, open it and check whether it supports the particular claim.
In practice
How much is a "token" in practice
In these examples, 1 token is roughly 4 characters in English. A word may take 1, 2–3 or more tokens, depending on the tokenizer. Usage includes the request, answer and file content used by the model.
← Swipe the table →
| Text | Tokens | Note |
|---|---|---|
| Hello there | 3 | ~4 chars = 1 token in English |
| Artificial intelligence is changing the company. | 8 | short office sentence |
| 1 page of A4 text | 500 | ≈ 500 tokens |
| Entire book (~300 pages) | 150,000 | illustrative example; check the particular model’s tokenizer and limit |