This is a follow up to this text ( https://laibyrinth.blogspot.com/2023/11/chatgpt-is-much-easier-to-use-than-most.html ), that aims to go more in-depth. and explain further details.
When news about ChatGPT spread around the world, I was, like many people, very curious, but also quite puzzled. What were the possibilities of these new ChatBot AIs? How did they work? How did one use them best? What were all the things they were "useful" for - what could they accomplish, and how?
My first "experiments" with ChatGPT often did not go so well. Add all this together, and I decided: 'I need further information'. So I looked online for clues and for help.
I quickly ran across concepts like "Prompt Engineering", and terms associated with it, like "Zero Shot Reactions". Prompt Engineering seemed to be the "big new thing"; there were literally hundred of blog posts, magazine features, instruction tutorials dedicated to it.
News magazines even ran stories which predicted that in the future, people who were apt at this 'skill' called "Prompt Engineering" could earn a lot of money.
And the more I read about it, and the more I learned about using ChatGPT at the same time, the more I realized what kind of bullshit concept prompt engineering and everything associated with it is.
I eventually decided to stop reading texts about it, so excuse me if I'm missing some important details, but from what I understand, "Prompt Engineering" means the following concept:
'Finding a way to get ChatGPT to do what you want. To accomplish a task in the way that you want, how you envision it. And, at best, using one, or a very low number of prompts.'
Now this "goal" seems to be actually quite idiotic.
Why?
Point 1 - Talk that talk
As I described in the text linked above (in the intro): ChatGPT is, amongst other things, a ChatBot and an Artificial Intelligence.
It was *literally* designed to be able to chat with humans. To have a talk, dialogue, conversation.
And therefore:
If you want to work on a project with ChatGPT, if you want to accomplish a task with it: Just chat with ChatGPT about it. Talk with it, hold a conversation, engage in a dialogue about it.
Just like you would with a human co-worker, collaborator, contracted specialist, whatever!
If a project manager wants an engineer that works for him to create an engine for an upcoming new car design, then he wouldn't try to instruct him just using 2-3 sentences (or a similar low number). He would talk with him, and explain everything, with as much as detail possible, and it would probably be a lengthy talk. And there would be many more conversations that follow as the car design project goes on.
So do the same when working with ChatGPT!
Obviously, companies try to reduce information noise and pointless talk, and reduce unnecessary communication between co-workers, bosses, and employees. But companies rarely try to reduce all their communication to "single prompts"!
It is unnecessary, and makes things more complicated then they should be.
Accomplish your tasks by simply chatting with ChatGPT about them.
Point 2 - Does somebody understand me? Anyone at all?
Another aspect behind the concept of "prompt engineering" seems to be:
"ChatGPT is a program with huge possibilities and capabilities. But *how* do you use it? How do you explain to ChatGPT exactly what you want?".
The "prompt engineer" then becomes a kind of intermediary between the human user and his visions of a project and his desired intentions, and the ChatBot AI. The user tells the "prompt engineer" his ideas and what he wants, and the engineer then "translates" this into a prompt that the AI can "understand", and the ChatBot then responds with the desired output.
But as I said above. There is no need for a translator or intermediary. You can explain everything to ChatGPT directly! You can talk to ChatGPT, and ChatGPT will understand you. Just talk to ChatGPT using "plain english" (or plain words), and ChatGPT will do the assigned task.
Point 3 - The Misunderstanding
This leads us to the next point. A common problem with ChatGPT is that while it understands you in terms of language, words, sentences, conversation, meaning - it sometimes still misunderstands the "project" you envision (partly, or even wholly).
This gives rise to strange output, false answers, the so-called "AI hallucinations".
Prompt engineering is supposed to "fix" this problem.
But it's not necessary!
If ChatGPT misunderstood something, gave "faulty" output, "hallucinates", and so on, then mention this to the AI and it will try correct it, and if it does not do that, keep talking.
Just like you would do in a project with human creators.
Example:
An art designer is told: "put this photograph of [person x]'s face to the background of an alien planet". The art designer does this. And then is told: "Oh, nice work, but we didn't mean an alien planet in the sense of H.R. Giger, but in the sense of the Avatar movie. Please redesign your artwork in that way." And so on.
Thus you need to work with ChatGPT in the same way.
True, sometimes this approach will not work (see below for the reasons). Just like not every project with human co-workers will get finished or be successful. But "prompt engineering" wont fix that either, then.
Point 4 - Shot caller
Connected to this is the case of "zero shot reactions".
I can understand that this topic has a vague scientific or academic interest, but literally zero real world use value.
"Zero shot reaction" means that an AI does the "right thing" after the first prompt, without further "prompts" or required learning.
But why would you want that? Sure, it takes a bit less work with your projects then, so if you're slightly lazy... but what use does it have above that?
Let's give this example: you take a teen that essentially knows zero things about basketball and has never played this sport in his life, and tell him to throw the ball through the hoop - from a 60 feet distance.
He does that at the first try (aka zero shot). This is impressive! No doubt about it.
But if he had accomplished that on the 3rd or 4th try, this would be slightly less, but still "hell of" impressive. Zero doubt about it!
Some might say the zero shot reaction shows how a specific AI is really good at understanding things; because it managed to understand the thing without further learning.
But understanding complicated matters after a few more sentences and "learning input" is still extremely impressive; both for a human and an AI.
This topic will be continued in part 2 of this text ( https://laibyrinth.blogspot.com/2023/11/forget-prompt-engineering-part-2.html )
No comments:
Post a Comment