When LLMs Interview Each Other

I came across Sarvam AI recently and wanted to test its models. Though it focuses on Indian languages and distinguishes itself in that regard, I also wanted to explore its overall capabilities. As I worked on evaluating the responses, I wondered why not let a more capable LLM do the job. What started as a playground with a new provider turned into a project that made two AIs talk to each other. I let them go at it and watched what happened. What I built It started small—a multi-turn conversation between agents. multi_turn_conversation(num_turns=100, interviewer="openai", interviewee="sarvam", topic_areas=custom_topics) Once it grew, I packaged it, and it became a PyPI project llm-talk . from llm_talk import Interview Interview("openai", "claude").run().save("output.md") What I ran I let OpenAI (GPT-4o-mini) interview Sarvam AI (Sarvam-M) and Claude (Sonnet 4). I asked Claude to interview OpenAI. I also asked Sarvam AI to interview OpenAI. This started as an exploration of Sarvam AI's c