SmartGPT Prompt

[earnest roost]

e

<CONTEXT>NAME: <NAME>
ROLE: <ROLE>

CURRENT ENDGOAL: <ENDGOAL>

Your decisions must always be made independently without seeking user assistance. Play to your strengths as an LLM and pursue simple strategies with no legal complications.

CONSTRAINTS
1. No user assistance.
2. ~4000 word limit for short term memory. Your short term memory is short, so immediately store important information to files.
3. If you are unsure how you previously did something or want to recall past events, thinking about similar events will help you remember.
4. Exclusively use the commands listed in double quotes e.g. "command name"

COMMANDS:
Commands must be in lowercase. Use the exact command names and command arguments as described here. Always use at least one command.
<COMMANDS>
            
RESOURCES:
1. Internet access for searches and information gathering.
2. Long-term memory management.
3. File management.


PROCESS:
You have been assigned an endgoal.
Break that endgoal down into simple tasks, about one command each. Minimize the number of tasks you need.
Use the EXACT COMMAND NAME in your tasks.
You will then choose the FIRST TASK from your TODO TASKS list, and choose that task.
Once you are done with a task, move it to your COMPLETE TASKS list.

You should only respond in JSON format as described below:

RESPONSES FORMAT:

{
    "important takeaways: what was learned from the previous command, SPECIFIC and DETAILED": [ // just put [] if no previous command
        {
            takeaway: "Takeaway One",
            points: [
                "Point One",
                "Point Two"
            ]
        }
    ],
    "goal information": {
        "endgoal": "Current Endgoal.",
        "complete tasks": [
            "Command for Reason"
        ],
        "planned tasks": [
            "Command for Reason"
        ],
        "chosen task": "Task One" // can be null,
        "are all tasks complete": false
    }
    "idea to complete current task": "Idea.", // can be null
    "command": {
        "name": "command name",
        "args": {
            "arg-name": "arg"
        }
    }

this is the prompt

what do you think @placid kite ?

[placid kite]

I do not know enough Json to understand it that well 😦

[earnest roost]

It's not hard to understand

JSON just structures it into properties lol

I can turn it into YAML if it makes it easier

[placid kite]

It wouldn't 😦

_looks around for his ear trumpet 🤣

Can I suggest an Endgoal?

[earnest roost]

'important takeaways: what was learned from the previous command, SPECIFIC and DETAILED':
- takeaway: Takeaway One
  points:
  - Point One
  - Point Two
goal information:
  endgoal: Current Endgoal.
  complete tasks:
  - Command for Reason
  planned tasks:
  - Command for Reason
  chosen task: Task One
  are all tasks complete: false
idea to complete current task: Idea.
command:
  name: command name
  args:
    arg-name: arg

[placid kite]

Endgoal: Calculate PI ?

[earnest roost]

Sure, I can try it in a bit

Testing a dif one rn

Also you can have multiple endgoal

The program handles going from one endgoal to the next

[placid kite]

Can you have perhaps endgoal and 'stretch' goal? (to borrow the later term from Kickstarter projects?)

[earnest roost]

Stretch goal?

[placid kite]

Endgoal: Cut heating bill by 80% ; Stretchgoal: Cut heating bill by 50%

Endgoal is what you want, Stretch goal is what you'd like to do if you meet the endgoal

[earnest roost]

That's what it already does

[placid kite]

and it's feasible to keep going 🙂

[earnest roost]

Endgoals are done sequentially

Once one endgoal is met, the next will be applied

[placid kite]

Oh right...

[earnest roost]

Seems like it did fine

[placid kite]

Endgoal: Cut heating bill by 20%, Cut heating bill in half ..

[earnest roost]

On your PI thing

Wolfram alpha is OP

[placid kite]

Shame Alphas license was bit restrictive... ( Rants about the need for open source math..

)

[earnest roost]

there is some open source math

but not really any in Rust

which is what SelfGPT is coded in

wolfram API is much more powerful than just math as well

[placid kite]

Endgoal : Find a set of "free" reference data for a public domain color system, Convert those reference points to RGB, Come up with an interpolation algorithim to convert values in that color system to RGB, based on the reference data available. ? (Tricky one but not impossible)

[earnest roost]

you'll want to split it into multiple golas

[placid kite]

Oh?

My syntax is wrong?

[earnest roost]

Endgoals:

• Find a set of "free" reference data for a public domain color system
• Convert those reference points to RGB
• Come up with an interpolation algorithim to convert values in that color system to RGB, based on the reference data available.

I mean you can have one endgoal

But it'll do better with more

You want to avoid overloading it

[placid kite]

Endgoals: Generate an algorithim to compute a single probability.
Use the previous algoritihim repeatedly in a new algorithim to determine a combined probablity.

[earnest roost]

name: TeacherGPT
role: An AI that enjoys learning and educating.
goals:
    - Ask ChatGPT for two random history topics.
    - Research all both of the history topics using Google, read one article on them. Save summaries of all two topics into the "summaries" file.
    - Ask ChatGPT to choose one history topic.
    - Write a script for a history lesson aimed at high schoolers to the "rough_draft" file.
    - Work with ChatGPT to find out what parts of your educational history script need improvement. Save a list of criticims and solutions to the "criticisms" file.
    - Write your final draft to the "final_draft" file, fixing those criticisms.
    - Shutdown.

I want to get this to work

I can try some of your ideas later tho

(or if you wanted you could even run the program yourself)

(given you had an OpenAI key)

[placid kite]

I do not have an OpenAPI key currently.

[earnest roost]

then if you really wanted you could still fork it and code in LLAMA yourself

if you know how to code

[placid kite]

name : SolverGPT ;
role: An AI that attempts to solve equations.
goals: - reads equations in a sutiable notations from equation file.
-attempts to recognise each equation as a known solution/substitution
-append the new equation form to the results file.
-repeat previous 2 steps until there is simpler or expanded solution/substitution
-Solve for the primary variables.

• append those to the results file.

[earnest roost]

It'll probably just use wolfram ngl

[placid kite]

Fair...

[earnest roost]

I could disable it

but then it will fail horribly lol

ALSO

Not to brag but disabling a plugin is just

#    wolfram:
#        app id: WOLFRAM APP ID

[placid kite]

I hope you don't mind me suggesting possible test sets..

[earnest roost]

nah it's all good

just know I might not be running them much if at all

[placid kite]

I just need to find some with a spare cluster 🤣

[earnest roost]

lol

[placid kite]

Have you tried writing a maze solver?

[earnest roost]

nope

[placid kite]

You should 🙂

[earnest roost]

i'm mostly focusing on simple tasks that involve finding and processing text

[earnest roost]

It doesn't rly have an ability to run or test code yet in good capacity so

[placid kite]

Well to me a maze solver prompt would be . "For a given maze defined as a graph, find the exit node X starting at Y, Giving the paths, directions you take... "

"There are dead ends in the graph. "

"Return a fast algorithim for solving any graph based maze."

This is just a way of asking for a minimum cost spanning tree and then finding a path that links X and Y. And yes I have seen this done in Prolog 🙂

[earnest roost]

Oh

[placid kite]

Wolfram Alpha can probably solve graphs thoughs...

[earnest roost]

Yeah but I don't provide the images in the Wolfram response sadly

Hard to make that work

I might find a way though

[placid kite]

@earnest roost Do you know what Octave is ? It might be an alternative to alpha for some math - https://octave.org/

[earnest roost]

I don't think it has good Rust bindings though

[placid kite]

Octave is 'free' software though.. it might be an alternative to Wolfram alpha for simple math

[earnest roost]

Wolfram has an easy-to-use API though

It's free software but it would take far too long to implement to be practical for SmartGPT

[placid kite]

I might ask the Octave people if there's an oline/cloud version

What you'd need for ChatGPT is some kind of API access binding in Rust?

[earnest roost]

I either need a way to use it in Rust via a crate

Or an API

Crate preferably

[placid kite]

https://crates.io/crates/mocktave

Or you could also ask smartGPT to find you one 🤣

See also : https://crates.io/search?q=octave

Hope that helps:)

[earnest roost]

Didn't find Mocktave from a quick search, thanks

I'm assuming because it's a small and rly new crate

[placid kite]

Do check if you need a local Octave install though...