The wrong words create the wrong projects.

Claude decides on its own how much to reason based on how complex the question is. You do not have to set anything.

An AI able to learn and reason like a human being, with no domain limits. It does not exist yet.

Autonomous systems that do not just answer: they have goals and memory, make decisions and carry out actions.

The techniques for making an AI follow human values and goals, avoiding harmful shortcuts.

One person per department (not necessarily from IT) who becomes autonomous with AI, works out what functions and spreads it to colleagues.

An AI system with a role, rules and quality standards that carries out a complete job on its own. Not an assistant waiting for instructions.

The mathematical structure trained for a task: the engine that generates the answers (GPT, Claude, Gemini).

A software layer that makes an AI model easy to use inside an app, without touching the API.

The company behind Claude, founded by former OpenAI researchers, focused on safe and reliable AI.

How two pieces of software talk to each other. It is the bridge between "AI exists" and "AI works in my job".

A visual output Claude creates inside the conversation: pages, charts, dashboards. Something you can see, use and export.

Standardised tests for comparing models. Useful, but they do not tell the whole story: one can shine in tests and disappoint in real use.

A technique where the AI reasons step by step instead of firing off a direct answer. It improves accuracy on complex problems.

Search across past conversations: Claude pulls relevant context from your old chats.

A program that simulates a conversation. It can follow fixed rules or be powered by an LLM that understands context and intent.

The conversational interface at claude.ai: you write, upload files, think. The entry point of the ecosystem.

The terminal agent: it writes code, manages repositories, runs scheduled tasks. Used by non-technical people too, for automation.

The agentic desktop environment: it accesses your files, installs plugins, runs tasks and creates documents. An AI office working for you.

The visual design tool: prototypes, decks, slides, mockups. It can hand the work to Claude Code to turn it into real code.

An Office add-in that reads the workbook cell by cell, changes values while preserving formulas, and fixes errors.

An Office add-in that respects the deck's formatting and creates content with native objects (real charts, not images).

A browser extension: Claude sees the page, navigates, fills in forms, extracts data. An assistant watching the screen with you.

Scans code for vulnerabilities and generates security patches. For cybersecurity teams and developers.

The instruction file Claude reads at the start of a Cowork session: the permanent operating briefing, specific to the project.

The shared knowledge layer every AI agent in the company reads and updates: procedures, decisions, templates, context.

The computing resources (GPU, CPU, memory) that run the models. More compute means more power and more cost.

The branch of AI that sees and interprets images and video: objects, faces, text, actions.

The bridge between Claude and your tools (Slack, Drive, Notion, GitHub). A Connector is an MCP already configured and ready in Cowork.

The information the AI holds in mind during a conversation. The more context you give, the more relevant the answers.

How much information Claude holds in mind in one conversation. The leading models reach 1 million tokens: like reading a whole book.

The Morfeus category: systems in production (data, decisions, automation, reliability) on a continuous basis, not an "AI agency".

A subfield of machine learning that uses neural networks with many layers. It underpins almost all modern AI.

The logic by which Claude decides which tool, skill or action to use for your request.

The representation of words or concepts as numerical vectors: similar concepts have nearby vectors. Used for searching and comparing.

The ability to understand why a model produced a given output. Crucial in healthcare, finance and justice.

When Claude reasons out loud before answering: better answers on complex problems, but slower.

Specialising an already-trained model with specific data, to adapt it to a domain or a style.

A large model trained on very diverse data, generic and adaptable (GPT, Claude, Gemini).

AI that creates new content (text, images, code, video) by predicting what should come next.

The rules you give Claude once (tone, format, context) and that it follows in every conversation.

A processor for parallel computation, born for video games and now the engine of AI. Training a large model takes thousands of GPUs.

Verified, reliable data used to train or validate a model. The reference point accuracy is measured against.

When AI confidently generates false content. It is not a bug but a characteristic of LLMs: it has to be managed.

The moment the model receives an input and produces an output. Every answer you get is an inference.

The date beyond which the model knows nothing, because it was trained on data up to that point (unless it searches the web live).

The time the AI takes to answer. Often it is the trade-off between quality and speed.

A model trained on enormous amounts of text to understand and generate language. It does not think: it predicts the best next word, but so well it looks like thinking.

The technique by which AI learns from data instead of being explicitly programmed. The heart of modern AI.

The proprietary Morfeus AI infrastructure, installed inside the company: it collects the data, automates, and improves with every project. It is not a SaaS.

The Morfeus frame: finding and recovering the margin a company loses every day while it scales. It is not cost cutting.

The highest level of Extended Thinking: no reasoning limit, for the hardest problems.

Model Context Protocol: an open standard created by Anthropic to connect an AI to external tools. It is how AI stops merely talking and starts acting.

MCP tools that return interactive interfaces (dashboards, forms), not just text, live on your real data.

The system by which Claude remembers information between one conversation and the next: preferences, role, projects.

An AI that understands and generates not only text, but also images, audio and video.

A structure inspired by the brain, made of connected nodes. It is the brain behind deep learning.

Natural Language Processing: the field that makes machines understand human language (summaries, sentiment, translation).

Open models (downloadable and modifiable, e.g. Llama) against closed ones (accessible via subscription or API, e.g. Claude, GPT).

The Morfeus model: an embedded AI team working inside the company with accountability for the results, not a per-project supplier.

The three Claude models: Opus the most powerful, Sonnet the best quality-speed compromise, Haiku the fastest and lightest.

The internal variables the model adjusts during training. More parameters means more capacity to catch nuance.

Packages that add skills, tools and connectors to Cowork in one click. Like apps, but for your AI environment.

Workspaces where you upload context files Claude reads before every conversation in that project.

The instruction you give the AI. Not a question: a delegation. The more precise it is, the better the result.

The craft of writing effective prompts. You do not need to code: you need to understand how AI reasons.

Retrieval-Augmented Generation: the AI searches real documents before answering. Fewer hallucinations, more answers anchored to facts.

A model designed to reason logically and sequentially, explaining its own process.

AI learns by trying, failing, and receiving rewards or penalties. Used for games, robotics, optimisation.

The Morfeus calculator that shows how much you lose every month and how much you can recover with AI.

The Morfeus AI layer that automates the sales team's repetitive work, giving hours back to the activities that generate revenue.

Tasks Claude runs automatically at set times, even when you are not there (Monday reports, competitor monitoring).

A file that teaches Claude a specific task. Like hiring a specialised employee in 5 seconds: you can create it, share it, install it.

Training on already-labelled data (input plus correct output), e.g. recognising spam from thousands of classified emails.

The invisible instructions that define who the AI is and how it behaves. The user does not see them, but they determine everything.

How much the AI gambles: low = precise and predictable, high = creative and unpredictable.

The process that breaks text into tokens (pieces) the model processes. Everything you write and receive is measured in tokens.

Claude's ability to use tools inside a conversation: search the web, read files, run code, call APIs.

A processor designed by Google to accelerate training, optimised for neural networks at large scale.

The phase where the model learns from data. It can take months and cost millions in compute.

The architecture underlying every modern LLM (2017). It handles relationships between words even far apart in the text.

AI learns from unlabelled data, discovering patterns and structures on its own.

The invisible margin losses in processes that companies do not see while they scale. Morfeus quantifies them in euros.

The monthly report with which Morfeus tells the CEO how much value it generated, in euros and not in slides.

Writing software by describing in words what you want and letting the AI generate the code. Possible for non-programmers too.

The internal numerical values the model learns during training and that determine its outputs. "Open weight" means they are public.