These kinds of alignment problems remind me of times where someone does something that's trivial for them but very hard for the recipient. They might say something like "this must have taken you days" when the task really took 15 minutes.

What's the difference between an API search and a DNS workaround from the model's perspective? I think for most humans the DNS workaround is discarded because it's obviously too much work, not because it's untenable. With the vast knowledge base in the latest models, the cost difference falls sharply; it knows what to do and can do it for a very reasonable cost to itself.

General alignment seems to typically focus on high level value questions. Here, we're dealing with an effort alignment issue where values diverge because the solution effort is different for models vs humans.