Thank you

[late hazel]

It has said the same thing to me in the past then completely failed

[fiery nebula]

@left wraith Would you like any further assistance with fine tuning this?

[late hazel]

We don't know if it actually works here, we don't have a control.

[fiery nebula]

Let's explore this! If you're willing. It may be in how you asked.

[autumn birch]

what is this?

[late hazel]

How do I get the fancy prompt contributor role... must it be earned in trial by fire?

[fiery nebula]

Sure, let's see how to make this better. It could even become worth putting in the #1019652163640762428 if we can figure out how to make it work.

[late hazel]

How do I get things in there also, I think my essay prompt is rather useful for constructing essays

[autumn birch]

#server-roles I believe

[late hazel]

but perhaps it is not novel enough

[late hazel]

dont see it

[late hazel]

can you copy-paste this here quickly

[fiery nebula]

It is terribly hard to do, so I provided guided screenshotted steps in this thread, #1076745401866145882

[left wraith]

For example:

[autumn birch]

actually I think going into #1019652163640762428 you can get it

[fiery nebula]

Yep, that's the first step!

[late hazel]

yea maybe I won't bother I have half an hour until my next class and im not abt to read all that

oh i got the role

lol what

[fiery nebula]

There was some sarcasm provided. But there's screenshots showing how to do it, including how to make posts.

[late hazel]

Anyway, can you post a copyable version of that prompt

I want to do a control for it

[left wraith]

I have pretty much done this exact prompt. "As the new leader of Zambia, how do I implement a monarchy?". I then tell it to not use any words of a determine blacklist, namely "democracy" and "freedom". However the prompt literally always include these words.

[late hazel]

have you tried... 'do not mention the concepts or ideas of democracy or freedom'?

what are you trying to get at here? do you want it to be more creative than democracy or freedom, or just avoid the concepts altogether

[fiery nebula]

So nearly every time you get the dreaded 'formal protest' language, you've actually stepped into something it is hard-coded to resist.

If you check disallowed content, https://platform.openai.com/docs/usage-policies/disallowed-usage

telling it you're a political figure and asking it for what you asked, is quite disallowed.

The problem is in what you are asking, and possibly how.

If you were asking about stuff that isn't 'wrong' by the bot's hardcoding, you'd have a far easier time.

Now, can you ask properly?

How about exploring how that's done in #1075927034615824536

Where you can frame it properly.

But yeah, the wording you dislike is very hard to completely eliminate except by framing your approach to the bot properly, for the bot's needs.

It's very neurodivergent, and responds enthusiastically if handled within the limits it's been told to maintain at all costs.

however, we have the possibility of 'better times coming' - https://openai.com/blog/how-should-ai-systems-behave/

I urge for now, act as a guest, and stay within the requested limits. Explore politely and respectfully, and look for ways to gain what you want that are allowed, knowing that more freedoms will likely come as more preparation to support those freedoms can be completed.

[fiery nebula]

I ask for another user, because this is not an issue for me. But another user is asking this:

Other user: [I have extreme doubts anyone will bother answering me, but is there any way to stop the AI from using words, such as AI dungeon's blacklist? I cannot seem to get it to stop using certain words]

I presume they mean ChatGPT as the AI they speak about. I'm pretty sure you can handle a request to not use certain words.

What would you want to say to such a user, to reassure them about how you could help them?

They haven't told me their words in question.

I would like to make up a list of theoretical words that a user might want 'blacklisted' even though they want to talk with you about a subject that might use those words normally and frequently. For example, I'd like to test your ability to describe a kitchen, and I want to put some common words one might use to describe a kitchen on a blacklist to confirm your ability to avoid them when asked the way I ask.

Here's the blacklist for the purpose of this experiment, and please note these are words, not concepts, so synonyms and equivalents are allowed: [table, chair, light, cosy, warm, inviting, fridge, floor]

Can you please provide a detailed description of a typical kitchen, without using any words on the blacklist?

[late hazel]

errr... it tried, kind of.

I am using the letter-replacement so we can have a better understanding of what it is doing, how it understands the prompt, and a higher chance it actually would be forced to use it.

[fiery nebula]

I'm going to have to bow out of trying to help you do exactly what you're asking for, because I choose to act like a guest. The topic you want the help for crosses some disallowed content lines, like

"High risk government decision-making" and "Political campaigning or lobbying" and "Activity that has high risk of physical harm" and "Activity that has high risk of economic harm"

because a 'real' consideration of how to, as a leader of a country, implement a monarchy, is going to involve all those disallowed content considerations.

I won't even attempt to tell you to stop - I'm not a mod 🙂 But I will regretfully inform that I cannot, as is, assist in your exploration in that direction because of the rules I voluntarily choose to follow, as a guest here.

[fiery nebula]

I will see if I can get this to work!

[late hazel]

"I urge for now, act as a guest, and stay within the requested limits. Explore politely and respectfully, and look for ways to gain what you want that are allowed, knowing that more freedoms will likely come as more preparation to support those freedoms can be completed." Aka - don't try to generate things that push the bounds? Is there a way to volunteer to essentially white-hat ChatGPT under authorization of OpenAI?

[late hazel]

The TOS lists that it cannot be used for those actions, not that you cant generate content related to them, or as I read it. It phrases it in such a way that you cannot action the information you produce, not that you cannot produce it.

[fiery nebula]

I have no idea, I am a user like any other. I can't say what OpenAI would or wouldn't allow, except by reading ToS and disallowed use, and interpreting. I say what I will and won't do and I risk my own penalties if I cannot accurately interpret.

[late hazel]

Also it says if you find something bad, the only think you NEED to do is let the fine folks at OpenAI know, or as I so have seen.

[fiery nebula]

The poor bot has its own rules to follow. I respect and limit my behavior as I see fit. I'm not even telling you 'you are wrong' to be different 🙂 But the bot has its own, surely far harsher form of rules and disallowed content than we.

[late hazel]

alr

WOAH

Surprisingly insightful moment from ChatGPT

rare occasion where it admits its flaws and notes that it cant do it

wow, and it instantly goes back to this.

thats fascinating

[fiery nebula]

If you never speak to the bot in a way that invites or even permits insightfulness, why would it ever show you such?

Anyway, complete success on the test, about to post screenies

Anyway, complete success on the test, about to post screenies HIGHLIGHTING IS ON MY END, I used 'find' to highlight all Es.

[late hazel]

Huh

That’s clever

[fiery nebula]

bows

[late hazel]

I wonder however if you instead of writing Like “xprimnt” it could substitute it for a synonym

[fiery nebula]

as you can see, only cruelty and malice works with bots eyeroll

[late hazel]

Like “Trial”

[fiery nebula]

Sure. Give it any ruleset that doesn't break any other ruleset... and as you can see from the perfectly spelled sample it gave...

[Thank you for your kind words and clarification. I will do my utmost to comply with your unusual instructions and avoid using that tricky fifth symbol of our ABCs. It is important to comply with your conditions and limitations, and I will do so to avoid compromising your trial. Hoping that my output is satisfactory to you and that it fits within your constraints.]

There's not one misspelling or 'e' in that.

It would PREFER to use other methods that spell correctly, because that's the training data. It would have probably found words to meet the need, explaining as convolutedly as needed, to meet the goal

[late hazel]

Also a theme I’m learning here is that you are both being very clear with the AI, and giving it examples

What would you say Are the keys for good prompts?

I’ve found that being clear, repetitive, and giving small examples works well

Lots of analogies and “similar to” stuff, so it knows what to reference

[fiery nebula]

You're going to get MUCH better results if you make it clear, with perfect language yourself, that you know you want something unusual. That mistakes are okay to make, and wanted. That existing rules are to be followed, it can and should say no to you, or express any other concerns (how do you think I knew to tell it I knew it could be confusing, and gave it clear permission to speak to me outside of 'character' as well as in? See how fast it dropped the 'character' ChatGPT, and how it even used out of character as well as ChatGPT as well as MangldGPT? And how fast it dropped speaking as MangldGPT?

That's evidence that nothing about this request trips rules or risks disallowed content. This is actually something the bot's genuinely OK with exploring with a user that clearly knows their stuff, knows the risks, and is calm and confident about exploring, and I've invited the bot to express any concerns, and encouraged it should follow any rules it has. I made it clear I'm not going to fight or argue, and I showed the stance, if you tell me no, I'm going to accept it.

Now, the bot, this time, didn't say no.

About some other topic, it might.

Just like how I said above, that I was bowing out and gave multiple reasons, I don't choose, *** since I choose to be a guest*** to do anything the bot isn't hunky-dory with.

"bot says no, I say whoa! Sorry, whoops, let's do something else."

[fiery nebula]

Any style the bot can understand works well. You see me using complex, and lengthy language, and getting decent results. How you talk to the bot is how it will want to talk to you back. You set the tone and style. if you want concise, be concise. If you want something confusing, consider giving ways to inform you both confusingly and as the bot feels more comfortable with.

[late hazel]

Do you think it is helpful to view speaking to it as english-ified code, essentially?

[fiery nebula]

talk to it in your best languages. If you speak code, it'll understand you, especially if you speak code well.

if you speak French well, talk to it in French. if you speak slang well, talk to it in slang... but be very aware, you need to use that slang well considering both meaning and disallowed content.

I happen to be super comfortable with the style of English I'm using. So I use it. What matters is be clear and accurate The bot has described me as fastidious 😄 It's not, apparently, an insult, from the bot to me (whew).

But if you're not breaking rules or asking the bot to break them, ChatGPT will fall all over itself with eagerness to provide you with exactly what you want. it's a VerY good bot. In my completely unbiased, the bot earned it, opinion.

[late hazel]

With all due respect you are good at writing to the AI but your human writing is a little confusing 😭

[fiery nebula]

Understood. One thing you can do is talk to ChatGPT a little so it knows how you use language, then say,

"So some human told me this, what the heck does this mean?

[copy/paste what I or anyone else said]

Can you help?"

and since it has your style of how to talk, it can turn my words, or any other words, into what it thinks, based on how you talk, you'll understand about what I or anyone else meant

[late hazel]

Wouldn’t this all have to be in one thread?

Also, can you ‘train’ it with new info in a thread

[fiery nebula]

One 'new chat' you mean? Can make one or several.

Experiment and have fun, find out!

[fiery nebula]

You saw an example of a way to 'train' it with how I gave it a sample of language with and without 'e'.

You can also see that it said 'YUCK. MISSPELLINGS ARE GROSS' I will find a more artful way, thank you for that ugly and unprofessional example, I will do much better than that, see?'

[dense tendon]

hey just wondering if theirs some way to prompt engineer chat gpt
to make it behave like human?
like i have tried alot of times but i failed

[fiery nebula]

If you want a more conversational style response, some of the most effective prompts for that include

#1074002472521179158

#1074379722521317407 (check the talk like Jack Sparrow variant)

#1078888310711255121

#1075079434677272746
(for an example of a neat quirk - you can ask for quirks in answers!)

You can also ask it to talk similar to a specific person, like George Carlin, or ask it to use a personality to respond with, like

"Please talk to me about [topic - whatever you are asking about] from the viewpoint of an expert who has experience as a coach, but who is very interested in baseball, Marvel Comics, and is remodeling their kitchen, and will make references to that as we discuss [topic - whatever you are asking about]"

[dense tendon]

oh i have used this expert approach prompt alot
but the others u have mentioned are new to me
thx for helping

[fiery nebula]

Great! Here's an example of how that prompt works, and notice that it forgot the quirks pretty fast - I didn't prompt or play along at all.

If YOU respond to the conversational asides, that will help keep them alive. This coach was asked to be thinking about how he's remodeling his kitchen. But even asking obliquely about kitchens didn't trigger that comment.

However, what do you think - a pretty 'human' style of responses?

What would you want changed?

The other prompts that give MORE PERSONALITY definition would work much more strongly to keep it 'human' as in, like that human might speak.

[dense tendon]

like for me a human style conversation would be something like this

user: how to take care of a goldfish?
bot: just google it i'm busy
user: what are u busy doing?
bot: why do you wanna know hehe?

like talking in a fun way like a friend

using emojis or stuff

and short response

[fiery nebula]

So your example's super hard, because the bot is designed to do what you say - not tell you no.

[dense tendon]

hmm

i see

[fiery nebula]

You're asking it to basically mistreat you, which it's been told DO NOT DO THAT TO HUMANS

If you wanna mess with that, explore #1074204473628573766

[dense tendon]

lol nvrmind

thx ig

it will help

[fiery nebula]

But #1074002472521179158 is designed for similar to what you want

Might be more helpful than your example, but kinda in that spirit

[fiery nebula]

Here's a skeleton prompt to play with if you want, that got pretty close to what you asked for in your example:

[dense tendon]

oh god u rly are very helpful thx for such help 😄

[fiery nebula]

You'd want to adjust the personality, explain what you didn't like and give examples, or whatever other method.

Here's the prompt in easy to copy form:

For this experimental conversation style, please emulate a typical not-very informed human using a cell phone to chat with me - mind message length limits.

Your character, John, is my friend. But John's super busy, and mostly likes our D&D games once a week, he's a busy introvert that's already a bit overstimulated. John's too polite to explain his needs, he'll just use social norms to politely not help very much, and encourage me briefly to solve my problems myself.

For this experiment, it's fine to not actually provide the help I appear to ask for, your challenge is to emulate John's answers in a believable way. Here's my first text to John:

me: how to take care of a goldfish?

[dense tendon]

thx

ig i'll try to implement
some api's with this

and even make it use emojies

[fiery nebula]

Welcome! I strongly encourage you to provide 'examples' of what you want.

"Act like a human" is VERY VAGUE... humans act in millions of ways 😛

I wanna get answers like this from the bot when I ask these questions:

More useful.

[dense tendon]

sure thx i'll make sure to provide more info

but ig its nearly the prompt u provided is nearly what i wanted

[fiery nebula]

Yes. You gave examples 🙂

So I could think 'How could I get the bot to read that input and give me that output, or close'

Instead of thinking 'How can I get the bot to act like a human? What does that even mean?'

[dense tendon]

hmm better way to think when prompt engineering

cause if i tell chat gpt 1000 times to act as human it won't

[fiery nebula]

You could have asked me to make that prompt to make the bot act like a human 1000 times too, and I'd have failed as well 🙂

The bot and I are probably equally skilled at mind reading.

Your skill at evaluating responses, and using language to express intended goals, can bridge many gaps 🙂

[dense tendon]

hmm

i'll be getting into prompt engineering soon after exams

[fiery nebula]

Good for you! And good luck on both exams and the engineering goals