The Reason Why Claude is Writing Less Humorous Content While Becoming More Intelligent

Jackson Kernion is an engineer at Anthropic and works on tuning Claude models. The reason why the latest Claude models generate duller text while appearing more intelligent? They’ve been optimized for code and math, explains Jackson. The explanations are optimized for other AIs, not for us humans. So a kind of jargon has emerged (“like a secret language among professionals”, as Jackson puts it himself). I found these kinds of texts annoying since Opus 4.6. Now with Opus 5.5 I think there’s a better balance, but it’s important to note that there’s nothing fundamentally different about this model.

I noticed something similar when using Claude Code: Often the explanations were quite technical and therefore sounded more like a report than a conversation. At times this was annoying because I wanted human readable text, not code. But for generating code this is perfectly fine of course. If you want to use the generated text for adding comments, communicating with customers or writing websites though, this jargon can be problematic. I hope Opus 5.5 has found a good middle ground between Opus 4.6 and the newest model. Not for coding, but for all other cases where the AI generates text that should be understood by humans.