Komplette Analyse des "Weltmodells": Definition, Ansätze, Praxis und ein weiterer Schritt in Richtung AGI
Das nächste Jahrzehnt der KI?
### Summary of the Text (World Models vs. Large Language Models)
The text discusses **world models**—AI systems that simulate the real world internally to predict future states, plan actions, and learn efficiently—comparing them to large language models (LLMs).
---
#### 1. What is a World Model?
A world model is an agent’s internal simulation of the real world, enabling it to:
- **Represent the world**: Understand objects, their positions, and relationships.
- **Predict the future**: Simulate events (e.g., a glass falling, a ball moving).
- **Plan and act**: Decide optimal actions based on predictions.
Its core components (from David Ha & Jürgen Schmidhuber’s 2018 work):
- **Vision (V)**: Extracts essential features from sensory input (e.g., table tennis ball movement).
- **Memory (M)**: Simulates future states using learned physical laws (like a "mental physics engine").
- **Controller (C)**: Learns optimal actions in the internal simulation (avoids real-world trial and error).
Example: A table tennis beginner uses V to capture ball features, M to simulate its path, and C to plan the best hit—all without repeated mistakes.
---
#### 2. World Models vs. LLMs
| Aspect | Large Language Models (LLMs) | World Models |
|-----------------------|-------------------------------------------------------|-----------------------------------------------|
| **Objective** | Predict the next word/token (linguistic coherence). | Predict the next state of the world (physical/dynamic changes). |
| **Training Data** | Mostly static text (plus images/videos). | Dynamic, time-ordered data (videos, robot sensor feedback, action results). |
| **Output** | Text/images. | Future state predictions, behavior simulations, actionable plans. |
| **Learning Method** | Indirect (via language, acting as a knowledge repository). | Direct (via interaction and simulation—can "see", "predict", and "act"). |
---
The text cuts off, but the key takeaway is: World models aim to make AI more "agent-like" (able to interact with the world) rather than just a "knowledge generator" (like LLMs).
(Note: The text has minor typos, e.g., "Weltmodelodellen" → "Weltmodellen", and is incomplete at the end.)