I built Andrej Karpathy's "LLM Council" on my own hardware, and now no single model gets the last word

Shekhar Vaidya rebuilt Andrej Karpathys LLM Council for local models on an RTX 4070 Ti GPU. The system runs DeepSeek-R1 8B, Qwen 3.5 9B, and Gemma 4 E4B via Ollama to debate prompts before a chairman model synthesizes the final answer. Sequential processing replaced parallel calls because simultaneous inference dropped responses on limited VRAM memory.