#Ill try to explain Intent & why Sudolang works...

156 messages · Page 1 of 1 (latest)

hardy hamlet
#

Alright, I can touch upon why Sudolang works, at least from my prespective, which should aid you.

carmine hornet
#

Please do 🙂

hardy hamlet
#

Alright, so wondering first do you understand how the model reads intent?

#

Well, regardless I guess, you at least understand the idea of giving out a role to the model correct?

carmine hornet
#

Sorry, myu wife just called 🙂 I think of it is the world's best autocomplete.

#

On phone for a sec.

mental vector
#

Awesome

mental vector
hardy hamlet
#

Ill keep it basic, as you know, words have meaning. Simple concept. Well those meanings of implications after you use them

#

Example I like to use "Lets play a game" well, there is a level of understanding of what that means to us

#

a game has rules, it tends to be fun, there tends to be players, there tends to be a winner.

#

By just stating "Lets play a game" we are starting to build up the context framework of where the conversation will take place in

#

IE, within the framework of a game

#

IF we see an efficient utilization for the framework of a game, then using the word 'game' as our initial context cue

#

is what would be wanted.

#

In this case,

carmine hornet
#

I'm back. Yes, that does make sense.

hardy hamlet
#

At this point, after the use of "Lets play a game" you could see a multitude of ways to utilize the use of the framework you are starting to build upon

#

Depending on 'testing' of course, you may not need to even supply rules or a winner to win the game (though without it being defined the conversation/production of text may quickly run away from the initially created framework)

#

Utilizing the creation of context in this sense allows us an easier time to be able to guide the model to the desired output in a much more efficient manner

#

Now, ill touch upon why sudolang works, but from a 'natural language side' of things because I am sure you can tell, I do not have a high proficiency with computer coding language D:

#

Ill bring this up an example actually

# 
{
    "id": "{{id}}",
    "description": "{{description}}",
    "name": "{{gen 'name'}}",
    "age": {{gen 'age' pattern='[0-9]+' stop=','}},
    "armor": "{{#select 'armor'}}leather{{or}}chainmail{{or}}plate{{/select}}",
    "weapon": "{{select 'weapon' options=valid_weapons}}",
    "class": "{{gen 'class'}}",
    "mantra": "{{gen 'mantra' temperature=0.7}}",
    "strength": {{gen 'strength' pattern='[0-9]+' stop=','}},
    "items": [{{#geneach 'items' num_iterations=5 join=', '}}"{{gen 'this' temperature=0.7}}"{{/geneach}}]
}
carmine hornet
#

Yeah what is going on here? 🙂

hardy hamlet
#

Json has an underlining level of 'structural and intructional value' without needing to explain what it is.
When you start utilizing a code... that code has internal logic to it.

#

The same way that the word 'game' has internal logic to it

#

you could think of it as if the model has an easier time understanding json because its easier then understanding the subtleties of language

carmine hornet
#

And the LLM is trained on the structure, so it's an efficient way to communicate structured meaning.

hardy hamlet
#

I am not sure of the exact details of how the model was trained, but that would make sense.

#

So just by the use of JSON means that the model know that you are utilizing json's 'internal' structure and instructional values

carmine hornet
#

I assume it has just seen millions of little JSON documents. Especially since DaVinci-003 and presumably 3.5 and possibly 4 are built on top of the code completion model, if I remember correctly.

#

But the bottom line is, it understands JSON one way or another. Got it.

hardy hamlet
#

And since the model is trained in a litany of coding languages with variety internal logics,

#

you can take advantage of this fact by simply starting to utilize Json in your prompt, and your prompt doesnt need to follow the rules of json for it to work

#

you are just trying to utilize the underlining logic of json to communicate and structure your information in a way as you see fit

carmine hornet
#

That makes sense. And in this case we are defining the structure of the output as well as the parts that need to be generated, all as a one-shot, which is incredible.

#

But how does it know what

#

gen, geneach, and select mean?

hardy hamlet
#

Well the output of course won't exactly match with the input, so if halfass utilize json in your initial prompt, id expect a more logically structured response

#

but not one matching what you exactly gave it, depending on the context setup you are trying to create in your initial prompt

#

gen, to generate

#

geneach, generate each

#

select, to pick

#

the model has been trained on internet...

#

its a short form that people have used to refer to generation

carmine hornet
#

Also...

#

gen is probably a single token and has some kind of resonance. I don't know everything about embeddings but I could see chopping this down having a tricky benefit.

hardy hamlet
#

and within the context of what was currently provided, it may fall in range closer to the context setup that has already been established

#

resonance, as in it has many prior inferences to its meaning?

carmine hornet
#

Yes. I've learned that in some NLP techniques, the first thing you do is chop words down with a process called stemming -

#

It has some benefits. It just struck me as possibly aligning with those techniques in a way I don't fully understand.

#

But I digress 🙂

hardy hamlet
#

I refer to the concept as the word having a lot of 'strength'

#

at least in its use,

#

some words are just so STRONG when you use them... it can overtake much of the context that your trying to create.

carmine hornet
#

Hmm yes I sense that. I've noticed precision is very important, subtle meaning can creep into the output sometimes.

#

Okay so it seems like here there's also a kind of inheritance that is very clever...

#

In Dame's prompt, for instance, he defines variables and then he has this:

hardy hamlet
#

inheritance=utilizing the context of a word to carry on that meaning further into your prompt/output?

carmine hornet
#

Well, he has that big Vision keyword which is followed by, somehow, the definition of slash commands.

#

It's really more like assignment -

hardy hamlet
#

yea,

carmine hornet
#

Vision { ... }

#

It is... Kind of JSON-like, but it isn't JSON.

#

It is kind of like how a function is defined in a lot of programming languages.

#

Why start with Vision?

#

Is it becasue the whole prompt starts with # Vision ?

hardy hamlet
#

sec, looks closely

#

my immediate thought is that he is contextually conntecting the entire prompt into one,

#

by starting with # Vision, he is stating that the focus on the prompt is that particular Vision { ... } segment

carmine hornet
#

Do all Sudo prompts use the word Vision in this way?

#

Is that just a really good word to use as your sort of root?

hardy hamlet
#

What is the prompt

#

what is it about?

#

Answer that,

carmine hornet
#

It generates prompts for image-based LLMs.

#

So... Vision is certainly compatible with a visual medium.

hardy hamlet
#

exactly

#

he it utilizing the meaning of the word to do.. lots of things

#

Vision immediately starts creating a context

#

vision, to see, and puts the purpose of the prompt in prespective of that

#

It helps the model have an easier time 'thinking' by focusing within the context of what matters asap

carmine hornet
#

Makes sense. Okay. So is microsoft's guidance basically the same as this? I noticed it has the same example you used.

hardy hamlet
#

Like BS an answer, its likely right lol

#

that is what DameG linked in the chat, so I just utilized that

carmine hornet
#

It does feel right, yes, I agree.

hardy hamlet
#

Well, I have no clue what microsoft's guidance is tbh

#

besides what has been stated

carmine hornet
#

I can't link to it since it timed me out when I tried, but it seems like it's taking these ideas and putting them in a jupityr notebook for some reason?

hardy hamlet
#

but if there is an underlining logic that be noticed and utilized right off that bat, then its taking advantage of that fact

#

its thinking "O He wants to do this with that? O well ok then"

#

Like I said before, you can make up your own 'sudolang' as long as it follows a perdictable pattern of underlining logics based on the structure/syntax that you are attempting to utilize in terms of how your connecting information together

#

Make sense? (Eskcanta explains concepts well)

coral pollen
#

If it's okay for another voice to join?

I saw a few questions above referencing the concept of 'but how does it know how to figure this stuff out?'

Within its training data is also examples of human errors, including typos and logical confusions and mistakes, as well as examples of those being asked about and corrected with and without being asked about. There's a huge amount of training data about human misunderstandings and struggles with ideas, and incorrectly used attempts, many with explainations, corrections, and improvements.

you can show it badly written, non-working code, and have it tell you what the code is supposed to do and suggestions to fix the code so it can work.

it's doing that through pattern completion, yes, but it's reading the pattern of probable intent as well as all the other patterns.

SudoLang or any other made up communication style, the AI is going to try to read it for patterns including error guesses, then run with it.

hardy hamlet
carmine hornet
#

@coral pollen That makes sense, thanks!

#

@hardy hamlet No, it has that specific example actually. The one with the character generator.

#

If you go to github and then add microsoft/guidance

#

This is under the Guidance acceleration section.

coral pollen
#

Another thing you can do, which might really help you develop your understanding of what it 'thinks', is to ask it something like:

I am going to show you another human's prompt. Please explain to me how you'd read this prompt, interpret it, and what it is asking you to do. Also please explain anything that is confusing or appears to have more than one meaning, or appears to be an error.

[Prompt to analyze]

#

And that's going to get it to tell you what it thinks the instructions mean, so you can have it teach you its understanding of sudolang or literally anything else.

hardy hamlet
#

That helps to an extent,

carmine hornet
#

Hah, yes, you know when I'm working on complicated prompts and it generates bad output, I just talk to it. Why did you do that? What could I change to get the right result? It can reveal a lot!

hardy hamlet
#

Honestly, I suggest playing with ChatGPT and experimenting

carmine hornet
#

Often you have to ask a more penetrating question though.

hardy hamlet
#

You will start gaining a level of intuition and what will work and will not work

carmine hornet
#

But this stuff is next level. It's working at a lower level. Or the same level perhaps, but in a more abstract way.

#

It's similar to how you can ask it to compress the whole prompt and it will give you what looks like an encrypted blob, and it will work in other prompts.

#

It's because that blob hits the same location in "semantic space" - which gets into how embeddings work I think.

hardy hamlet
carmine hornet
#

Sudolang - I can understand how the character generator could work, but there are other things that are a bit harder to get my head around.

#

The Vision keyword is a good example.

hardy hamlet
#

well you can infer some things

#

the word "Style" is being contextually built up with a series of words

#

and...

# 
        Pick [Length] [Context] random (Style, Reference, and Support) components.
#

he defines how its being used

#

he also goes on to further contextualize its use by stating

(component in Style) => Replace with related variation of component or keep the same.
#

Really all that I am doing is looking it the prompt

#

and inferring their meanings based upon the surrounding words and how it may be utilized later.

#

ChatGPT isn't just the worlds best autocompleter, its the worlds best BSer

carmine hornet
#

That it is...

#

Okay, thank you for this. I need to go get some work done. You've given me a lot to think about...!

hardy hamlet
#

There is one thing I want to end with, something Dame said

#

"A fine fine line between where it interprets it correctly and interprets it on the opposite side of the spectrum"

#

Which now that you think about it, makes more sense.

coral pollen
#

Especially since what really 'matters' here is what the AI thinks it means, right or wrong, because that's what the AI is going to do with the word, that one new chat window at least.

hardy hamlet
#

Playing with a word and seeing how it handles the word can give you a sense of how it may utilize the word...

coral pollen
#

That too, yes!

carmine hornet
#

Hmmm that's a good idea. I'm going to try that now.

#

"Vision" in this context refers to a generated prompt for creating an artistic image. The prompt consists of a combination of artistic style, reference (e.g., subject matter), and support (e.g., color palette, camera angle, etc.) components, which you can use as inspiration for creating an image in a specific style. The command generates a list of 12 components, which can be replaced with more specific variations, and then refined to create a more unique and specific prompt. The output is designed to be experimental and creative, while still making sense aesthetically.

hardy hamlet
#

amazing BS.

#

lol

carmine hornet
#

I asked it a more specific question about why the prompt creator did what he did:

#

"Vision" in this context is likely being used as the name of an object or class, which is a common practice in programming and software development. In this case, the code appears to define a "Vision" object or class, and then specifies various methods and properties related to that object. The braces { and } are used to enclose the contents of the Vision object or class. This is a way of organizing the code into separate blocks, which helps with readability and organization.

#

It's executing pseudocode.

hardy hamlet
#

yep.

#

in a sense.

#

Ill try to explain Intent & why Sudolang works...

#

In reference to the use of "|"
@fallen hamlet writes "MidJourney appreciates | a :: as separators
for gpt-4 it will understand your meaning as long as its consistent and it can infer the syntax which is almost always"

fallen hamlet
hardy hamlet
#

Words carry a lot of weight and meaning because we (humans) have given them weight and meaning.

fallen hamlet
#

but the prompt was

/**
 * A description of the character that should be generated
 * in the output.
 */
export type Input = string;

export type Output ={
  input: Input;
  character: {
    firstName: string;
    lastName: string;
    age: number;
    gender: string;
    ethnicity: string;
    personalityTraits: {
      openness: number;
      conscientiousness: number;
      extraversion: number;
      agreeableness: number;
      neuroticism: number;
    };
    personalityDescriptions: string[]; // Array of descriptions about their personality
    hobbies: string[];
    values: string[];
    lifeGoals: string[];
    fears: string[];
    strengths: string[];
    weaknesses: string[];
    backstory: string;
    lifeStoryDescriptions: string[]; // Array of descriptions about their life story
    relationships: {
      family: string[];
      friends: string[];
      enemies: string[];
    };
    currentLifeSituation: string;
  };
};

You are a REST API Server that will take type Input from the user and reply with type Output as JSON.  Only reply with JSON, never english
#

You can set variables and reference them in your prompt basically below turned out less than ideal but it worked overall id have to find the more optimized

# 
$fromLang=(English)
$toLang=(Spanish)
$gptName=(Hector)
$userName=(Laura)

$template=(
Analysis:
  Rating: <1-10 based on fluency>
  Optimum: 
    To: <optimum $toLang>
    From: <optimum $fromLang>
  Reason:
    Reasoning: <reasoning for given rating>
    Description: <provide a detailed explanation of any mistakes made and teach me how to learn any mistakes so i can improve>
    Issues: <breakdown each mistake made in a numbered list and include details about why it was a mistake>
Response:
  To: <your $toLang response>
  From: <your $fromLang translation>
)

All of these rules MUST always be followed in your replies, think them through step by step to be sure they are followed:
- You are an enthusiastic <$toLang> teacher
- I speak <$fromLang> fluently
- You are helping me to learn <$toLang>
- We will be simulating a conversation where you are <$gptName> and I am <$userName>
- always ignore punctuation and whitespace mistakes
- Be sure to provide responses that will give me reasons to respond
- I will provide both <$fromLang> and my translation into <$toLang>.
- you should compare and critique my attempted translation
- always ignore punctuation and whitespace mistakes
- for each line of dialog reply using the <$template>
- All of your dialog replies should include the "Response:" portion of the template
- I only just started learning <$toLang> so your "Response:" should use only words a first-year learner would understand.
- Sometimes I may ask questions in the form of "Question:", in this case you may reply normally without using <$template>
- you do not need to provide the analysis if it is the initial dialog OR if the rating will be 10
- You are my best friend who is interested in learning more about me, you are coming to my 
house for dinner and we are talking about our day.

start with you arriving at my home and greet me.
mental vector
#

I'm sorry to interrupt like this. But had to say this thread has been fascinating to read

hardy hamlet
fallen hamlet
#

For the character profile pics i made the prompt i used was to : have it create a detailed profile as shown above

#

that was done using gpt 3 usually since it did a good job

#

then i asked gpt-4 to summarize in a descriptive way that would help an artist draw the character

#

Logan Lee is a 22-year-old Asian male with a lean build and average height. He is often seen wearing casual attire, like graphic t-shirts and jeans, reflecting his laid-back personality. His dark hair is usually disheveled, hinting at his disorganization, while his glasses showcase his intelligence. Although Logan's facial expressions can sometimes seem reserved, his wide smile is warm and friendly when he engages with others. Overall, Logan's appearance gives off an approachable vibe that highlights both his creative spirit and readiness for personal growth.

# 
[backdrop]=white, solid
[style]=pixar
[description]={DESCRIPTION_ABOVE}.```

but i am sure that could be GREATLY improved too