[onion2k]

In this case the LLM is just acting as a super-charged search engine.

It isn't, because that implies getting everything necessary in a single action, as if there are high quality webpages that give a good answer to each prompt. There aren't. At the very least Claude must be searching, evaluating the results, and collating the data in finds from multiple results into a single cohesive response. There could be some agentic actions that cause it to perform further searches if it doesn't evaluate the data to a sufficiently high quality response.

"It's just a super-charged search engine" ignores a lot of nuance about the difference between LLMs and search engines.

[hnaccount_rng]

I think we are pretty much past the "LLMs are useless" phase, right? But I think "super-charged search engine" is a reasonably well fitting description. Like a search engine, it provides its user with information. Yes, it is (in a crude simplified description) better at that. Both in terms of completeness (you get a more "thoughtful" follow up) as well as in finding what you are looking for when you are not yet speaking the language.

But that's not what OP was contesting. The statement "$LLM is _doing_ $STUFF in the real world" is far less correct than the characterisation as "super-charged search engine". Because - at least as far as I'm aware - every real-world interaction had required consent from humans. This story including