This could be said about any new tool.

"photoshop has created a dichotomy --- those who are below-average in their skills will see it as an improvement, while those who are above average will see it as regression."

"the automobile...

"the camera...

"the printing press...

signed: someone who really does not like these LLM's. Not for what they can do, but for the doubt they bring to the table. This is probably the same feeling as when digital image manipulation got good and you could no longer trust photos.(not that you ever really could. photos, even unaltered ones are notoriously easy to show only what you want them to.)

No, it couldn’t. The tools to which you compare require a different or complementary set of skills to use, and the user must understand the output or mode of operation that the tool conveys.

Prompt engineering (within the context of ChatGPT) is more useful for jailbreaking than using it for its intended purpose, aside from which, the complaint of the OP from the linked post is that their coworkers do not understand the output, if they did, they would not rely on ChatGPT as a smokescreen.

What about googling then? I often google how to do X while writing code or developing because I don't remember the exact CLI flags and reading the manpage takes more time than asking google. I'm not a big LLM user yet but I imagine a lot of people have replaced googling with ChatGPT in their workflow.

As someone who grew up around artists who eventually used photoshop when it was version 1 I can only express how completely ahistorical a recounting is for me. I also know a fair bit about the history of the camera and can firmly say that's also not how that went down.

Why would you just say this?

Assuming AI is at the level of average

AI has created a new standard from which “average” is defined. Rather than a real life analog to a mathematic mean, it’s simply understood to be the level at which AI performs.

It is a tool, those who learn to use the tool will benefit more. average people will be average in using it, smart people have the potential to use it smartly.

It naturally assumes the average of its training data.

I'm not sure this is necessarily the case, especially considering techniques like RLHF that can improve a model's performance.

This statement smells whole lot like Dunning–Kruger.