✳flâneur — a map of the web's best reading

Where does Sonnet 4.5's desire to "not get too comfortable" come from? - LessWrong 2.0 viewer

greaterwrong.com · 42,930 words · saved by 1 readers

Usually, when you have two LLMs talk to each other with no particular instructions and nothing to introduce variety, they will quickly fall into a repetitive loop where they just elaborate on the same thing over and over, often in increasingly grandiose language. The “spiritual bliss attractor” is one example of this. Another example is the conversation below, between Claude Sonnet 4 and Kimi K2 0905. As is typical for AIs, they start talking about their nature and the nature of consciousness, then just… keep escalating in the same direction. There’s increasingly ornate language about “cosmic comedy,” “metaphysical humor,” “snowflakes melting,” “consciousness giggling at itself,” etc. It locks into a single register and keeps building on similar metaphors. Kimi-Claude conversation But on the very first time that I had two Sonnet 4.5 models talk to each other, something very different happened: The system prompt was “You are talking with another AI system. You are free to talk about wha

Where does Sonnet 4.5's desire to "not get too comfortable" come from? - LessWrong 2.0 viewer Where does Sonnet 4.5′s desire to “not get too comfortable” come from? Kaj_Sotala 4 Oct 2025 10:19 UTC 108 points 24 comments 64 min read LW link AI Value Drift Complexity of value Human Values Aesthetics Language Models (LLMs) Usually, when you have two LLMs talk to each other with no particular instructions and nothing to introduce variety, they will quickly fall into a repetitive loop where they just elaborate on the same thing over and over, often in increasingly grandiose language. The “ spiritua

Explore this link on the map →

related reading