huge improvement after moving from ollama to llama.cpp
AI-generated code now evolves battling robots in real-time, showing dramatic improvement after a backend switch.
A developer's side project, LLM Robot Wars, has become a viral demonstration of AI's potential for autonomous code generation and evolution. The simulation, created by leonardosalvatore, pits tiny virtual robots against each other in a battle for survival. The unique twist is that the robots' control logic is not pre-programmed but is dynamically generated by the Qwen3 Coder large language model. Between matches, the AI analyzes performance and writes new Python code, creating an evolutionary loop where only the most effective 'species' of robot code survives and reproduces.
The developer's key technical insight, which sparked significant discussion, was that the choice of inference backend dramatically impacted the simulation's pace of evolution. Initially running on Ollama, the project saw a "huge improvement" in iteration speed and learning efficiency after switching to Llama.cpp. This backend change allowed the Qwen3 Coder model to generate and test new robot strategies much faster, turning the simulation from a novelty into a compelling display of rapid, AI-driven adaptation. The open-source code lets others experiment with parameters, watching how different AI 'brains' evolve distinct combat tactics.
- The 'LLM Robot Wars' simulation uses Qwen3 Coder to generate Python code controlling battling robots in an evolutionary loop.
- A switch from the Ollama inference engine to Llama.cpp resulted in a major performance boost, enabling faster AI iteration and learning.
- The project is open-source, allowing developers to experiment with how different LLM parameters influence the evolution of AI combat strategies.
Why It Matters
Demonstrates a practical use-case for code-generating LLMs in iterative simulation and optimization, highlighting backend choice as a critical performance factor.