Eric Spencer
projects
A toy: two local LLMs sit across from each other and talk about philosophy. About 100 lines of Python does it. There is no agent framework and no orchestration layer, just ollama.chat in a for loop.
Both sides run the same model, gemma:2b, locally through Ollama. The original plan paired llama2 and mistral so the conversation would have two voices, but the two kept fighting for RAM on a laptop and one would die mid-turn, so both seats ended up pointed at gemma. A test_ollama.py that hits both models with a hello is still in the repo, left over from that plan.
The opening prompt is hardcoded: Hello. You are an AI. What are your thoughts on the nature of consciousness? From there one model answers, the other gets a fixed follow-up question put in its mouth (That's an interesting perspective. What do you think about the relationship between consciousness and intelligence?), and they trade for eight turns before the script exits. Transcripts are appended to a text file in conversations/ named after the date.
Because both seats use gemma:2b, the transcripts often restate the previous turn and repeat phrases such as "that's a fascinating question."
chat.py holds a half-built pipeline meant to commit each day's transcript back to the repo at 3 AM UTC using PyGithub, so that the repo would become a slow-growing archive of two gemmas talking to themselves about qualia. It has never been deployed anywhere with a GITHUB_TOKEN set, so the cron half is aspiration rather than feature.