The article includes tables with visual representations of the random walks at various temperatures for different models, highlighting the aesthetic and random-like walks of the llama3:8b model. The author raises questions about why some LLMs cannot produce a random walk at a temperature of 0 and what might be causing the gemma2:9b model's peculiar behavior. The experiment was conducted using a Mac M2 with 16 GB RAM, and the results were color-coded for different configurations. The article concludes with links to additional resources, including code, animated frames, and videos of the random walks.
Key takeaways:
- The experiment involves testing LLMs like llama3.1/2 and gemma2 series to perform random walks, with an unexpected behavior observed in the gemma2:9b model.
- The gemma2:9b model fails to consider the UP and DOWN directions in the random walk, unlike other models which follow the instructions correctly.
- The experiment setup includes using Ollama with LiteLLM on a Mac M2 with 16 GB RAM, where each interaction is new per turn per walk without context continuation.
- The results are visualized in tables showing random walks at different temperatures, highlighting the aesthetic/random-like walks of llama3:8b and the peculiar behavior of gemma2:9b.