I've noticed this more and more recently, especially with Opus 4.6. Not super frequently, but enough that it stands out given how capable 4.6 is. Has anyone else seen this? Any theories as to what causes it?
There's nothing particular about this. This is just what you'd expect from a Language Model trained on large datasets. It reproduces a pattern commonly found in documents
Not sure I agree with you. For lower ability models, yes. Claude Opus 4.6 is incredibly capable, so it's odd to me it has this residual 'misspeak' behaviour.
That's the issue, people are anthropomorphizing those models, but... they're all the same (conceptually). They just do random hallucinations, trying to make those hallucination match the "reality" (of their training data) as much as possible
Maybe RL? Just like similar corrections in reasoning traces. You can train non-'thinking' models the same way (though if you're naive about it then you might end up with responses that are similarly rambly), and I'd expect it to have been
