morev•o
Glossary


Kein Treffer für diesen Begriff.
01 The Architecture of the “Brain” Models & Mathematics Fundamentals of LLMs, transformer technology, neural networks, and inference processes. AGI, Attention Mechanism, Backpropagation, Embedding, Latent Space, Parameter, Transformer 02 The Agent's Hands Frameworks & Tools Autonomous workflows, browser control, and practical task execution. Agentic Workflow, Auto-GPT, Code Interpreter, Context Window, OpenClaw, RAG, System Message 03 Security & Infrastructure Network & Protection Securing local and cloud systems, encryption, and access control. API, Containerization (Docker), E2EE, Firewall, Least Privilege, Prompt Injection, Sandbox, VPS 04 Data Processing & Knowledge Management The Memory Storage and preparation of information for AI use. Data Lake, Data Sovereignty, NER (Entity Recognition), Structured Data, Vector Database 05 Ethics, Society & Regulation The Guardrails Laws, societal impacts, and the “alignment” of AI. AI Act, Alignment Problem, Bias, Deepfake, Explainable AI (XAI), Turing Test 06 Advanced Technology & Programming The Gearbox Software environments, development tools, and configuration standards. Daemon, Environment Variables, JSON, Latency, Middleware, Node.js, Webhooks 07 Cybersecurity & Attack Vectors The Digital Firewall Threat analysis and defense strategies. Adversarial Attack, CVE, Exploit, Honeypot, Red Teaming, Zero-Day Exploit 08 Mathematical Mechanics & AI Training The Training The statistical processes behind machine learning. Batch Normalization, Convergence, Epoch, Gradient Descent, Hyperparameter, Overfitting 09 Autonomous Economy & Future Visions The Horizon Long-term trends, multi-agent systems, and the path to superintelligence. Autonomous Economy, Digital Twin, Emergent Behavior, MAS, Recursive Self-Improvement, Singularity 10 Industry Standards, Protocols & Niches The Expert Arsenal Highly specialized protocols and cutting-edge optimization techniques. GRPO, MQTT, REST API, Quantization (GGUF), Shadow IT, Zero-Knowledge Proof
01

The Architecture of the “Brain”

Models & Mathematics

AGI (Artificial General Intelligence)
The theoretical stage of AI where a machine can perform any intellectual task that a human can. Unlike today's “narrow AI,” an AGI would possess true understanding and transferable skills across all domains.
Attention Mechanism
The core of transformer architecture. It allows AI to focus selectively on the most relevant words in a sentence (e.g., linking a pronoun to its referent), enabling modern language comprehension.
Backpropagation
The fundamental learning algorithm. It calculates the AI's error after a response and sends this information “backward” through the network to adjust internal weights, reducing future errors.
Base Model
An AI model trained on vast raw data but without specialized “behavior.” It functions as a powerful text completer and serves as the foundation for later assistant models.
Cosine Similarity
A mathematical measure of similarity between two vectors. In AI, it quickly determines how closely two sentences' meanings align.
Deep Learning
A subset of machine learning using deep neural networks with many layers. It enables machines to autonomously recognize complex patterns in images, sound, and text.
Embedding
The conversion of words into high-dimensional numerical vectors. These allow AI to mathematically represent semantics: similar concepts cluster closely in “vector space.”
Fine-Tuning
The process of retraining a general model with domain-specific data (e.g., medicine or law). This transforms a generalist into a specialized expert for a niche.
GPT (Generative Pre-trained Transformer)
The architecture family that sparked the current AI boom. It uses transformer technology to generate text autoregressively (word-by-word) based on probabilities.
Hallucination
A phenomenon where AI confidently fabricates facts. This occurs because the model predicts statistically plausible words without real-world grounding.
Inference
The actual operation of AI. While training takes months, inference is the moment when you ask a question and the trained model computes the answer.
Latent Space
An invisible coordinate system where AI organizes all learned concepts. The AI navigates this space to associate ideas and create new content.
LLM (Large Language Model)
A massive language model with billions of parameters. It serves as the “operating system” for AI agents, unifying logic, language, and world knowledge in one system.
LoRA (Low-Rank Adaptation)
A highly efficient fine-tuning technique. Instead of retraining the entire model, tiny adapters are added, enabling training on standard consumer hardware.
Loss Function
The “teacher” during training. This mathematical function measures how wrong the AI is; training aims to drive this value toward zero.
Mixture of Experts (MoE)
A design where a model is divided into specialized sub-units. For each query, only the relevant “expert” activates, saving computational power and speeding responses.
Multimodality
An AI's ability to combine different senses. A multimodal agent can simultaneously read text, analyze images, and understand voice commands.
Neural Network
A computational system inspired by the brain. It consists of layers of artificial neurons that process signals through weighted connections, learning patterns autonomously.
Parameter
The “knobs” inside AI. More parameters (e.g., 70 billion) allow finer nuances and greater knowledge storage capacity.
Quantization
A compression technique for AI models. By reducing mathematical precision, models shrink in size and speed up, fitting onto laptops or phones.
RLHF (Reinforcement Learning from Human Feedback)
The “upbringing” of AI by humans. Testers rate responses, teaching the AI to be polite, safe, and helpful while adhering to social norms.
Softmax Function
The final mathematical layer that converts internal calculations into clear probabilities (0-100%) for the next word.
Temperature
The “creativity dial.” Low temperature makes AI factual and rigid; high temperature makes it imaginative but error-prone.
Token
The currency of AI. Text is broken into small pieces (tokens); the number processed determines cost and memory limits.
Transformer
The revolutionary 2017 architecture. It allows AI to consider the full context of a sentence simultaneously, rather than reading word-by-word from left to right.
Universal Approximation Theorem
A proof stating that neural networks can compute any logical function—they are theoretically capable of any conceivable cognitive task.
↑ Zum Inhaltsverzeichnis
02

The Agent's Hands

Frameworks & Tools

Agentic Workflow
A process where AI doesn't just respond once but works in loops: draft, self-critique, revise, then output.
Auto-GPT
One of the first autonomous systems. It uses LLMs to create its own to-do lists and execute them online or on local storage.
Browser Automation
An agent's ability to control a web browser like a human (clicking, scrolling, typing) to complete tasks on websites.
CLI (Command Line Interface)
Text-based control (terminal). Experts use it to launch and configure agent frameworks like OpenClaw precisely.
Code Interpreter
A tool that lets AI write and execute real code. It solves mathematical problems or analyzes data flawlessly.
Context Window
The “short-term memory.” It limits how much information (e.g., 100 pages of text) an agent can process simultaneously.
Headless Browser
An invisible browser without windows. Agents use it to rapidly collect data from websites in the background (scraping).
Human-in-the-loop (HITL)
A safety principle. The agent asks for human permission before critical actions (e.g., transferring money).
Moltbook
A social network exclusively for AI agents. Here, AIs exchange information autonomously, with humans allowed only as observers.
OpenClaw (formerly Clawdbot)
The leading framework for private AI agents. It enables operation on personal hardware with full data sovereignty.
Playwright
A technical tool for browser control. It serves as the agent's mechanical hand to perform complex interactions on web pages.
RAG (Retrieval-Augmented Generation)
The most important technique against hallucinations. The agent first checks your documents and uses these facts as the basis for its response.
System Message
The agent's “DNA.” It defines who the agent is, how it behaves, and which rules it must never break.
↑ Zum Inhaltsverzeichnis
03

Security & Infrastructure

Network & Protection

Adversarial Attack
A manipulation attempt where AI is deceived by specially crafted data (e.g., invisible pixels in an image).
API (Application Programming Interface)
The digital bridge. Agents “talk” to other programs (email providers, calendars) via APIs.
Containerization (Docker)
Packaging an agent into an isolated box. This ensures consistent performance across devices and prevents unauthorized system access.
E2EE (End-to-End Encryption)
Encryption ensuring only you and your agent can read data—no hackers or providers in between.
Firewall
The digital doorman. It controls who from the internet can access your agent's dashboard or commands.
Least Privilege
The golden security rule: grant your agent only the permissions it absolutely needs. An email bot doesn't need access to your bank data.
Prompt Injection
The biggest threat to agents. An attacker hides commands in text the agent reads, hijacking its control.
Sandbox
An isolated playground. Agents perform risky tasks here, so errors or viruses can't harm your real computer.
Tailscale
A VPN tool creating a secure “pipe” from your phone to your home agent, anywhere in the world.
VPS (Virtual Private Server)
A rented server online. The perfect place to run an agent 24/7 without interruption.
↑ Zum Inhaltsverzeichnis
04

Data Processing & Knowledge Management

The Memory

Data Augmentation
Artificially expanding datasets by slightly altering existing data (e.g., rotating images or adding noise). This teaches AI to be robust against variations.
Data Lake
A massive repository for raw data in its original format. Unlike structured databases, it stores unsorted texts, images, and videos, making them accessible to AI agents later.
Data Mining
Systematically searching large datasets to uncover unknown patterns, trends, or correlations. Agents use this to extract key business figures or market trends from thousands of documents.
Data Sovereignty
The right and technical ability to maintain full control over your data. With local agents like OpenClaw, this is central—no data leaks to external corporations; everything stays on your hardware.
Entity Recognition (NER)
An AI's ability to automatically identify specific information (names, places, organizations, dates) in text. Agents use this to create calendar appointments from email floods.
Knowledge Cutoff
The point when a model's training ended. Anything happening afterward is unknown to the AI unless it uses tools like web search to update its knowledge.
Metadata
Additional information about data (e.g., creation date, author, file size). Agents use metadata to efficiently sort information and grasp a file's context faster.
Structured Data
Data organized in a fixed format (e.g., Excel tables or SQL databases). AI processes it far more easily and accurately than unstructured text or audio.
Vector Database
The agent's specialized long-term memory. It stores information as mathematical coordinates, allowing the agent to search for semantically similar content in milliseconds.
↑ Zum Inhaltsverzeichnis
05

Ethics, Society & Regulation

The Guardrails

AI Act
The world's first comprehensive AI law (EU), classifying applications by risk. It bans dangerous practices (like social scoring) and enforces strict transparency for powerful AI models.
Alignment Problem
The existential challenge of ensuring an AI's goals perfectly match human intentions. A misaligned agent might pursue a goal in harmful ways.
Anthropomorphism
The human tendency to attribute feelings or consciousness to AI. This often leads to overestimating an agent's capabilities, as users forget it merely calculates statistical probabilities.
Bias
Systematic errors in AI, usually from biased training data. This can cause agents to discriminate against certain groups or amplify prejudices.
Constitution (AI Constitution)
A safety approach where AI is given a fixed set of moral rules. The model uses this “constitution” to continuously check its responses for ethical correctness.
Deepfake
Deceptively real AI-generated media (video, audio, images). They pose a disinformation risk, as it becomes harder for humans to distinguish real evidence from artificial fakes.
Explainable AI (XAI)
A research field aiming to make AI decisions understandable to humans. XAI tools explain why an agent made a specific decision.
Singularity
The theoretical point where AI growth becomes irreversible, surpassing human intelligence so profoundly that civilization's future becomes unpredictable.
Turing Test
A classic test for machine intelligence. If a human can't distinguish between a machine and a human in conversation, the test is passed.
↑ Zum Inhaltsverzeichnis
06

Advanced Technology & Programming

The Gearbox

API Key
A digital security key acting as an ID. Your agent uses this key to authenticate with services (like OpenAI or Anthropic) and access their computing power.
Containerization (Docker)
A technique to run software in isolated environments (containers). This prevents an agent from damaging your OS if it crashes and makes installation effortless on any machine.
Daemon
A background program running invisibly, waiting for tasks. The OpenClaw daemon ensures your agent can receive messages even when the program window is closed.
Environment Variables
OS variables storing sensitive data like passwords or API keys, so they're not visible in the agent's source code.
Git / Versioning
A system recording every code change. Developers can revert to a previous, working version if an update causes issues.
JSON
The standard data format for AI communication. It's human-readable text that computers efficiently convert into commands.
Latency
The delay between your input and the AI's response. Low latency is crucial for smooth conversations and fast agent automations.
Middleware
Invisible mediator software between applications. It helps agents prepare data from legacy databases so modern language models can understand it.
Node.js
The runtime environment for many agent frameworks. It allows using the web language JavaScript directly on PCs or servers for AI tasks.
Webhooks
A mechanism where an app instantly notifies your agent when an event occurs, instead of the agent constantly checking for updates.
↑ Zum Inhaltsverzeichnis
07

Cybersecurity & Attack Vectors

The Digital Firewall

Adversarial Attack
A targeted manipulation attempt where AI is deceived by subtly altered input data (e.g., making it misclassify a stop sign as a yield sign).
Brute-Force Attack
A primitive but effective method where an attacker systematically tries all possible password combinations. Modern agents defend against this with “rate limiting,” restricting attempts per time period.
CVE (Common Vulnerabilities and Exposures)
An international standard for cataloging known security vulnerabilities. Users should monitor CVE lists to know when critical patches are needed.
Exploit
A specific program or code snippet exploiting a known vulnerability. Attackers use exploits to gain control over agents or escalate privileges on host systems.
Honeypot
A deliberately vulnerable system designed to attract attackers. It helps study hacker or bot methods without risking production systems.
Phishing
Attempts to obtain sensitive credentials or API keys via fake messages (emails, SMS). AI agents are increasingly trained to detect such fraudulent intent automatically.
Prompt Injection
The most dangerous LLM attack: an attacker hides instructions in text the agent processes, tricking it into ignoring system rules and revealing private data.
Red Teaming
A simulated attack by security experts to test an AI's defenses. The goal is to find vulnerabilities before real attackers do.
Zero-Day Exploit
An attack targeting a vulnerability unknown to the vendor. With no existing patch, this poses the highest risk to autonomous systems.
↑ Zum Inhaltsverzeichnis
08

Mathematical Mechanics & AI Training

The Training

Batch Normalization
A technical procedure during training that stabilizes input values within the neural network. This accelerates learning and reduces sensitivity to poor initial weights.
Catastrophic Forgetting
The problem where a neural network, when learning a new task, completely overwrites knowledge of an old task. Researchers use special techniques to “freeze” important knowledge.
Convergence
The point in training where the AI stops making significant progress, as the error (loss) reaches a minimum. A well-converged model is ready for deployment (inference).
Epoch
One complete pass of the entire training dataset through the neural network. Most models require hundreds of epochs to internalize complex relationships.
Gradient Descent
The mathematical “hiker” searching for the lowest point in a foggy valley. It calculates how to adjust neural weights to minimize the AI's error.
Hyperparameter
Settings manually defined before training (e.g., learning rate). They determine the learning process's framework and are often decisive for success or failure.
Overfitting
An error where the AI memorizes training data. The model performs perfectly on known data but fails completely with new, unseen information.
Parameter
The billions of numerical values inside a model representing connection strengths between neurons. They are the actual storage units of learned knowledge.
Softmax
The final layer converting internal calculations into clear probabilities. For example: “I am 98% certain this image shows a cat.”
↑ Zum Inhaltsverzeichnis
09

Autonomous Economy & Future Visions

The Horizon

Autonomous Economy
A scenario where AI agents act as independent economic participants. They negotiate contracts, purchase computing power, and pay for services among themselves using digital currencies.
Digital Twin
A virtual replica of a real-world object (e.g., a factory). Agents use these twins to simulate scenarios risk-free before implementing physical changes.
Emergent Behavior
Abilities that suddenly appear in a model at a certain scale (e.g., logical reasoning), even though they were never explicitly trained. This makes large AI development fascinating and unpredictable.
Multi-Agent System (MAS)
A team of specialized AIs. Instead of one AI doing everything, a “swarm” collaborates, with each agent (e.g., coder, designer, tester) contributing its strengths.
OODA Loop
A military decision cycle (Observe, Orient, Decide, Act) serving as a model for autonomous agents. It describes how a system must respond to environmental stimuli in milliseconds to act effectively.
Recursive Self-Improvement
The theory that an AI could rewrite its own code to become smarter. This could trigger a rapid intelligence explosion beyond human comprehension.
Technological Singularity
The theoretical point where technological progress becomes so rapid that human civilization changes fundamentally and irreversibly.
↑ Zum Inhaltsverzeichnis
10

Industry Standards, Protocols & Niches

The Expert Arsenal

GRPO (Group Relative Policy Optimization)
A highly efficient optimization method known from models like DeepSeek-R1. It enables AI to train complex logical reasoning (reasoning) without a separate reward model, as in classic RLHF.
MQTT (Message Queuing Telemetry Transport)
An extremely lightweight network protocol for the Internet of Things (IoT). AI agents use MQTT to communicate with sensors and smart home devices, ensuring reliable data transfer even on unstable connections.
Long-Context Window
The ability of modern models to hold massive amounts of data (up to 2 million tokens) in “working memory” simultaneously. This allows an agent to analyze hundreds of PDFs in a single query.
REST API (Representational State Transfer)
The global standard for web interfaces. Almost every AI agent communicates via REST APIs with services like Google, Spotify, or payment systems to execute real-world actions.
YAML (Yet Another Markup Language)
A text-based format for software configuration. YAML is almost as easy for humans to read as a list and is used in agents like OpenClaw to securely define behavior rules and API keys.
Vector Clock
An algorithm in distributed systems that helps determine the temporal order of events across different agents. This prevents conflicts when multiple agents work on the same document simultaneously.
Headless Browser
A web browser without a graphical interface. Agents use it to navigate websites, fill forms, and collect data in the background without a visible window.
Middleware
Invisible software layer mediating between applications. It helps agents translate data from legacy systems into formats modern language models can process.
NPM (Node Package Manager)
The essential tool for managing extensions in JavaScript-based agent frameworks. Libraries for browser control, encryption, or database connectivity are installed with a single command via NPM.
Quantization (GGUF/EXL2)
Special file formats for compressed AI models. They shrink a model that originally required 100GB of GPU memory to run on a gaming laptop with 16GB RAM.
Shadow IT
The use of AI agents in companies without IT department knowledge. This poses high security risks, as sensitive corporate data could leak via private agents to the cloud.
Zero-Knowledge Proof
A cryptographic method where the agent proves it knows information without revealing the information itself. This protects privacy in automated logins and transactions.
↑ Zum Inhaltsverzeichnis