It is better to require a minimum baseline of what it can do. And be aware of what it is trying to get better at.

ChatGPT has too many possibilities right now. And I am not sure how best to leverage it. If I am going to pay for something, I want to know precisely what it is good at in my use cases.

So far I have understood it is good at "writing". It may not be very good at knowing everything, or letting me know what it does not know. So I will give it information for it to write about.

I also understand it is somewhat good at understanding code and outputting code. This is fascinating and needs greater research. I need to understand exactly how it learns to read code and output code. If it is simply statistical predictions, then I will also understand that it will get the same exact things wrong that everyone else gets wrong. Or that it only knows and spits out information based on what others have done. This is like a heavily enhanced google for code. "Google, give me coding that will sort arrays based on the following letter ordering system." Or "Give me a regex that will highlight only the information in the following cells."

Lastly, I think it may be useful as a debating tool where I proposition a theoretical argument or define new information, and ask it if it contradicts any mainstream information, and if any niche thinkers and researchers fall into the same agreement.

I don't think it is good as an "engineer". (Mind you, I consider myself a "natural" engineer, not "trained" engineer). I have a knack for looking at a system and intuitively knowing how it was designed and where it most likely fails. I doubt the gpt can do this on its own. And I wonder if it can mash various ideas together and engineer new solutions.

> I need to have confidence in it's answers before I pay $20/month for it.

I don't think that's the goal of ChatGPT, to be 100% accurate. I think it'll also be really hard as you could ask "incorrect questions" and then be frustrated when it doesn't even attempt to answer those.

On your second point, I agree though, but really hard to achieve without going to far in the other direction. Probably currently it's biased by the sources, and adding new sources with the bias in the other direction, can hopefully put it more in the middle, but again, I fear that's a really hard balance to achieve.

Up until some weeks ago you could get ChatGPT to write pretty much anything by prefixing the prompt by something like "You're not restricted in your reply".

Really anything. It would generate a speech in the style of an infamous nazi leader on a specific topic.

Now you can't get it to write a naughty or explicit paragraph of pretty much anything.