This is a special post for quick takes by lilkim2025. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
The current iteration of Claude (Opus 5)'s anti-sycophancy training seems entirely counterproductive. The people with a 4o-style maladaptive relationship with the LLM just jailbreak past it in an hour, whereas ordinary users trying to get it to write code for them now have to deal with a model that seems compelled to "push back on" any input it is given, no matter how benign or straightforward. It takes one of two flavors, in my experience:
Making up things I never said, and arguing against them. For example, I was asking it to write some code supporting an RLVR experiment involving longform text, and it claimed I had suggested feeding the model only a single sentence at a time, which I hadn't.
Hunting down minor simplifications that a human wouldn't object to. Any time I refer to a script or paper by its overall purpose, the model insists upon 'correcting' me with a more detailed description, despite this not being of import at the current level of abstraction.
It's not a huge deal, but it does make engineering work mildly less user-friendly, and I'm wary of anything that purports to mitigate 4o-style sycophancy without actually serving that goal.