LLM Telephone Game (Week 1)

Semantic drift experiment: pass one message through a chain of LLMs and measure how meaning degrades.

Week 1 foundations contribution — multi-model calling + LLM-as-judge + embedding similarity.

Chain

GPT-4o → Claude → Gemini → DeepSeek → Mixtral/Groq → Llama (Ollama local)

Features

• Universal API wrapper (OpenAI, Anthropic, Google, Groq/DeepSeek, Ollama)
• Strict "repeater" system prompts
• Quantitative score: cosine similarity (SentenceTransformers)
• Qualitative score: GPT-4o judge (mutation / hallucination)
• Matplotlib drift chart

Setup

cd 1_foundations/community_contributions/Sama-ndari_llm-telephone-game
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Create .env (do not commit):

OPENAI_API_KEY=...
ANTHROPIC_API_KEY=...
GOOGLE_API_KEY=...
# optional: GROQ / DeepSeek / local Ollama

Optional: run Ollama locally for the last hop (ollama pull llama3).

Run

Open llm_telephone_game.ipynb and run all cells.

Author

Sama-ndari — https://www.samandari.dev
Related repo: https://github.com/Sama-ndari/llm-semantic-drift-analysis