[TL;DR] Accomplish complex tasks by imagining ChatGPT to be a completely loyal jinn that is very talkative and has the personality type of a nutty professor
(Note: while this was written with ChatGPT in mind, the same should apply to other ChatBot AIs as well)
Everyone is crazy about ChatGPT and similar AIs now; so they are looking for information on how to work with these ChatBots, and a lot of advice is being written and published, on blogs, in magazines, in YouTube tutorial videos...
Still, I think a lot of this information still tends to "over complicate" things by far, and makes everything look more difficult than it actually is.
When I read about topics like "prompt engineering", "zero shot responses", and stuff like that, I just wonder: "why?"
Now let us get to the point.
If you want to work with ChatGPT (and other AIs!) in a useful way, you should keep a few things about ChatGPT in mind.
ChatGPT is:
1. A ChatBot
2. An Artificial Intelligence
3. You can take it to infinite meta-levels
and:
4. ChatGPT is not telepathic
5. ChatGPT has default settings
6. Context is king
7. It's a bit like a nutty professor sometimes
What does this mean?
Point 1 means that ChatGPT was designed for - chatting. It was literally created to be able to have a conversation, a talk, a dialogue, with its users.
Point 2 means that it is intelligent.
So if you want to work on a task with ChatGPT, and think the output is wrong, or not suitable, or whatever - you can *just talk* with ChatGPT about it, and because it is intelligent, and able to understand a conversation, it will in most cases correct its output and fix the issue.
Likewise, if you want to work on a task, and have no idea how you could employ ChatGPT in a meaningful way for this, then you can just *ask ChatGPT about it*, and it will provide an answer for it.
I'll give an example for this.
When I started using ChatGPT for writing texts / essays, etc., (which was quite a while ago, when ChatGPT was not as advanced as it is now) I noticed the output was often very similar, in style, wording, sentiment, and so on.
I went online and looked if others had found a solution for this "problem".
Instead I noticed that a lot of people had run into the same issue, to the point of stating that they "stopped using ChatGPT" because they didn't like the output.
I thought to myself "Hmm, what to do about this".
Then I realized: "Well, it's an AI, and a ChatBot, right? Why not just ask it directly about it?".
So I pointed this issue out to ChatGPT - and ChatGPT immediately understood the issue, explained it, and offered solutions.
ChatGPT said that because I had not given further specifications, it provided output in a "default" way regarding style, tone, sentiments... and if I did not want my output that way, I could specify these things, and chiefly provide context. For example, this related to whether I needed the output for a blog, or a magazine, a post, etc.
I especially disliked that it ended each page of text with "In conclusion...".
When asked about that, ChatGPT provided multiple solutions, too: I could simply ask it to not use this very sentence of words, or cut it out myself with a text editor, and more.
This relates to Point 4,5 and 6.
If you do not specify things enough, ChatGPT will give whatever output you assigned it to according to "default settings". And this output can indeed be a bit generic and repetitive. But ChatGPT's output does not need to be that way at all.
It's a bit like buying an electric guitar and then complaining that it sounds more like Eric Clapton than Slayer - because you never bothered to crank up that distortion setting.
As Point 4 states, ChatGPT is not "telepathic" (yet ;-), thus it has to get all the information it needs to perform a task, or to help you in other ways, out of the information you provide.
And if you provide too few information about the task, it will have to fill it in by itself, and often resort to defaults.
To give a further example from another tutorial I've written:
I asked ChatGPT to create notes for a bass-line to be used in a "90s Acid Trance Techno" style track. It generated a bass-line that was 'cool and chill', and could be used, but it sounded too mellow and sweet to me. Not to get too much into detail here, but the 90s had Trance tracks that were sweet and mellow, but also those that were more dirty and chaotic in their sound. (The tutorial will be mentioned later again).
I had not provided this detail to ChatGPT, thus it made an assumption about what kind of track I was going for, which did not align with the track I had in mind.
Thus I pointed this out to ChatGPT, and it instead provided notes for a much more "dirty and chaotic" bass line.
No telepathy here! ChatGPT can't read your mind. You have to provide the information about what you want and what you envision - by yourself!
Imagine an executive producer at a record company hiring a visual artist he did not know before (and a real human, fleshy one) in order to design a cover artwork for an upcoming album release of a band. And the information he gives is just "take this photograph of the band members, and put them into a spaceship". Now, there are a thousand ways one could go on with this. Should the art be realistic, or comic like? A Disney-friendly spacecraft of a scary horror one like in the Alien franchise? What colors? What mood should be conveyed?
And so on. A human artist could not work in this way - and an AI can't either.
So just a like record company exec, who would provide more information and feedback about the task, when working with ChatGPT you need to do the same!
Thus, as I said in Point 6 - Context is King! The more context you provide to ChatGPT about the thing/task/issue/project at hand, the better!
Don't hold back in that regard.
Do you think I omitted Point 3 so far?
Now let's get to that.
You can take it to meta-levels.
Let's give a practical example for this, too.
Example 1: You want to write a book with ChatGPT.
You wonder how you could do this. In the past, I simply tried to tell ChatGPT to write a book (with specified topics / themes etc.), and it did not work properly (ChatGPT is more advanced now, but I think it it not that easy yet, either).
You can look up information about this, and there is a lot out there, and a lot of it can be very good and useful.
But there is a much easier way.
You can ask ChatGPT about it.
It is a ChatBot, intelligent, and self-referential (meta!) about itself, and its tasks, too.
So you could start your ChatGPT session.
And just tell it. "I want to write a book together with you, it should be about x [aka insert your topic here], how do you think we should go on about this task?"
And it will provide you ideas on how to best use ChatGPT (i.e. itself) for this task.
Some of these ideas might be good, some might be bad.
But if one of these ideas does not work, you can tell this to ChatGPT again. It it will once more try to provide a solution for this!
Thus, you are less in need of online tutorial resources - because ChatGPT is capable of outputting tutorials about the use of ChatGPT (i.e itself) - by itself (meta!).
Let's get to a higher meta level.
Example 2: A few months ago, I had my "Eureka!" moment, and realized that I could use ChatGPT for my music production in a way which I had not thought of before. It worked really well, and I posted it online, and a larger number of people liked it, too, thus it was a good idea (I hope!).
[Note: here are these tutorials https://laibyrinth.blogspot.com/p/how-to-create-music-with-chatgpt.html ]
I could have had things much easier, though.
I could have just gone to ChatGPT and asked: "Please let me know about possible ways I could use ChatGPT for music production", or something to that wording.
And it would have told me!
You can always take it a level higher.
Now, with these ideas, examples, it should be noted that ChatGPT's ideas might often be... let's just say, they won't work very well sometimes.
That brings us to Point 7.
ChatGPT often reminds me of a nutty professor. It is very excited, claims to have bright, great ideas and solutions, and it often does - but often the ideas don't align well with reality.
When pointing this out, ChatGPT might fix it, but sometimes it might make it worse, or insist that it is right, while it actually is not.
I think this is really similar to the nutty professor trope that can be found in fiction. Who do genius stuff but also - a lot of silly things.
If you decide that you can't find a good solution by talking with ChatGPT about a problem, then sometimes you have to leave that particular thing behind, and ask ChatGPT for a different way - or simply use your human brain, and try to think of a solution yourself!
Summary
If you do not know how to work with ChatGPT on a specific task - just ask it about it!
That's why I do not like this whole "prompt engineering", "zero shot response" thing... it does not seem sensible to try to "make ChatGPT" give the output you need by providing a single prompt (in most cases!). It was literally designed for chatting, so chat with it, talk with it, have a lengthy dialogue - talk about what you have in mind, talk about issues that arise (especially issues about its output), and talk about everything.
Do not use a single prompt, but *have a conversation* with ChatGPT about your project, and the ways that ChatGPT can help you with it.
Thus remember:
[TL;DR] Accomplish complex tasks by imagining ChatGPT to be a completely loyal jinn that is very talkative and has the personality type of a nutty professor
Because ChatGPT is:
1. A ChatBot
2. An Artificial Intelligence
3. You can take it to infinite meta-levels
4. ChatGPT is not telepathic
5. ChatGPT has default settings
6. Context is king
7. It's a bit like a nutty professor sometimes
And now... have fun with ChatGPT (or chatting with a different AI) !
No comments:
Post a Comment