What is ZorkGPT?
ZorkGPT is an AI agent system that plays the classic interactive fiction game "Zork" using Large Language Models (LLMs). It demonstrates how AI can understand and navigate complex text-based game worlds.
What is Zork?
Zork is a legendary text adventure game first released in 1977, where players explore the Great Underground Empire by typing commands in natural language. Often considered the most famous piece of interactive fiction, Zork challenges players to solve puzzles, collect treasures, and navigate hundreds of locations using only text descriptions and typed commands.
Why Zork instead of Pokemon?
I've seen the impressive Claude and Gemini plays Pokemon projects and wanted to explore what was possible with open source models. Why didn't I try making open source models play Pokemon? Because I'm not as smart as the guys who created those projects! 😉 I originally tried to limit myself to models that could reasonably run on home systems configured for LLM inference, making this more accessible to hobbyists and researchers without massive compute budgets. However after thousands of turns, I've needed up expand to larger (still open source) models to get the best results.
What am I looking at?
This live viewer shows real-time data from a running ZorkGPT session, including:
- Game Info: Current episode, turn count, score, and location
- Knowledge Base: AI-generated strategic guides and learned patterns
- Game Log: Recent actions, AI reasoning, and game responses
- Map: Visual representation of discovered locations and connections
How does the AI work?
The system uses multiple AI components, each potentially running different open source models:
- Agent: Decides what actions to take
- Extractor: Parses game text into structured data
- Critic: Evaluates proposed actions before execution
- Knowledge Base: Analyzes gameplay data to generate strategic guides and learns from experience
Current Models:
- Agent: Loading...
- Critic: Loading...
- Extractor:
- Knowledge Base: Loading...
These models may change as experiments progress to test different capabilities and performance characteristics.
Why isn't the data updating?
The viewer polls for updates every 3 seconds. If you see a red connection indicator, either:
- No ZorkGPT session is currently running
- The state file isn't being generated
- The data is more than 10 minutes old (stale session)
- There's a network connectivity issue
What do the log colors mean?
- 💭 Yellow: AI thinking/reasoning
- 🎯 Light Blue: Proposed action (final action taken)
- 🚫 Orange: Rejected actions (if any) with critic justifications
- ⚖️ Orange: Final critic evaluation with detailed justification
- ✅ Light Green: Action taken
- 🎮 Green: Game response
What does the critic evaluation show?
The critic evaluation displays:
- Status: Whether the action was approved or rejected
- Score: Numerical evaluation (positive = good, negative = bad)
- Justification: Detailed reasoning for the evaluation
- Override status: Whether a rejected action was overridden
Can I interact with the game?
No, this is a read-only viewer. The AI is playing autonomously. You can watch its decision-making process and see how it learns and adapts.