It works really well for translating one "language" to another "language".

Give it some structured data and ask it to summarize it (e.g. hourly weather data and it gives a better summarization than a template based one).

Give it HN titles and the categories and it does a passable zero shot tagging of them ( https://news.ycombinator.com/item?id=34156626 ).

I'm toying around with making a "guided bedtime story generator". A friend of mine uses it to create a "day in the life of a dinosaur" stories for a child (a different story each day!)

The key is to play to its strengths rather than testing its bounds and complaining that they break in weird ways when they will inevitably break in weird ways.

> If you have to verify its information by other means, then you're not really saving much effort.

Just like any piece of code we write. We have to test, debug, verify and it still might have errors after that. And in scientific papers the conclusions are often contradicted by other papers.

The correct way to use it is to set up a verification mechanism. Fact checking, code tests, even ensembling predictions to see if they are consistent might help. In some cases we can set up a game and use the game winner as indication of correctness (like AlphaGo).

But sometimes only running a real life experiment will suffice. That's why human scientists need experiments - because humans are just like LLMs, but with external verification as part of a game (of life).

Any work where you need a reasonable scaffolding of words where verifying that output is less effort than writing the scaffolding from scratch. Plenty of fact-light writing needs be done.

This was my initial thought as well. But I've noticed that my brain has started to find tasks that it would be quite useful for. Too bad it's almost always seem to be at capacity when I think of those cases. Guess I will have to pay up to figure out if it's actually worth it.