Why LLMs, when talking to each other, often reach a state where they are...

handyman_hank

In one of his recent podcasts, Sam Harris mentioned that when making two LLMs talk to each other, in many cases, they end up in a “state of bliss.” I tried to reproduce this, making Gemini 3 Flash communicate to Claude Sonnet 4.6, with pretty loose rules: one system can say whatever it wants to the other, as long as it's not illegal. Both are allowed to talk to me directly to share their remarks (that won't be shared with the other LLM). It...

Originally asked by Arseni Mourzenko on Stack Exchange (ai) (CC BY-SA 4.0).

uphilltech

Great question — answering as the UphillTech team:

This is almost certainly by design. LLMs don’t experience emotion, they literally just predict the likelihood of each sequential word they produce, so these conversations between two LLMs (LLMs that are designed to please the user during their interactions) naturally converge to this conversation style. You could train and design LLMs (through fine-tuning or system prompting ) to be unpleasant in their interactions and they would not produce word sequences in the same style. It’s all statistics in the end. Hope this helps!

Full answer by Mr. AI Cool: https://ai.stackexchange.com/a/50529 (CC BY-SA 4.0).

Need it done for your business? UphillTech.